Explore the 26 original 1080p / 30 FPS stage videos →
What started as a simple prompt and a schematic browser Three.js telemetry ribbon evolved—through 75 commits, three cloud GPU architectures, and ten iterative engineering chapters—into a survey-grounded, 80 Hz physics-coupled photoreal rally stage renderer. Here is the complete story of how we built it, every dead-end we hit along the way, and the frame-accurate evidence behind what runs today.
We did not start with a game engine, a pre-built rally track, or a commercial asset pack. We started with raw
10 Hz World Rally Championship telemetry for #69 Kalle Rovanperä and #33 Elfyn Evans, a browser
window, and a question: could we conversationally build (“vibe-code”) an interactive 3D WRC visualizer from
scratch?
Each time the prototype worked, the next visual or physical limitation became immediately obvious in the frame.
Fixing those bottlenecks one by one forced us to outgrow the browser, bring up bare-metal cloud GPUs
(AMD Radeon Pro V710 → NVIDIA A100 → NVIDIA RTX PRO 6000 Blackwell), write custom
shaders for 260,000 hair-ribbon grass strands and two-phase hydrodynamic spray, subdivide the car to 682,182 triangles,
replace synthetic sine-wave scenery with national LiDAR/topographic surveys, and solve 80 Hz nonlinear vehicle dynamics.
Synchronized 10 Hz WRC telemetry and an A100 CUDA
friction-circle optimal ghost (#00) with 30 Hz University of Tartu RTK-GNSS trajectories, Estonian Maa-amet
1 m numeric LiDAR DTMs/nDSMs, Riegl point clouds, ETAK topographic vectors, and Finnish SYKE / Metsäkeskus forest inventories across
SS2/SS5 Elva, SS3/SS6 Neeruti, SS4/SS7 Otepää and SS22 Ruuhimäki.
When browser WebGL hit its ceiling, we moved rendering to cloud GPUs streaming 1080p frames back to the browser HUD. The renderer combines a 260,000-strand Cubic-Bezier hair grass system, an 8-species 4.10M-triangle scanned botanical forest, crown-first 3D Perlin-Worley cumulus clouds, two-phase gravel dust and Weber-breakup puddle spray, and corridor-relative golden-hour lighting.
The monolithic car GLB was topologically
split into five articulated assemblies (body + FL/FR/RL/RR), subdivided to 682,182 curved PN-triangles with
native 8K PBR livery and 12-pass 3D rotational wheel blur, and driven by an 80 Hz sub-stepped single-track bicycle ODE
fusing onboard accx/accy G-sensors, a +0.78 m front-axle drift pivot, parabolic crest flight, and
decoupled unsprung wheel knuckles.
Every major visual claim on this page is backed by matched before/after captures, 1080p/30 FPS motion clips, and explicit root-cause diagnostics—including the dead-ends we had to throw away (“bare brown sticks” tree imports, crushed ink-black shadows, faceted far-canopy sheets, and weightless turntable yaw filters).
What this is—and what it is not: This is an engineering and visual-computing journey showing how far iterative human-AI co-engineering can push a custom scientific/motorsport visualizer when every critique is answered with math, real survey data, and GPU code. It is not an official historical survey or a claim of 30 FPS live throughput at maximum offline capture settings; what is measured, modeled, or approximated is stated explicitly throughout.
How did we get from a question—“could we vibe-code a WRC racing visualizer?”—and a schematic
Three.js WebGL telemetry ribbon to the cloud-rendered digital twin shown above? Across 75 commits and ten major
engineering iterations (Chapters 00–09), every leap in visual and physical fidelity came from looking honestly at
what felt fake in the current build, diagnosing the exact mathematical or hardware root cause, and re-engineering the
subsystem—including several dead-ends that had to be openly abandoned. Click any image below to inspect it
in the full-screen lightbox (use ← and → to step chronologically across the
entire evolution, or jump to the Vehicle & Mechanical Feature Zooms and the
80 Hz Vehicle Physics A/B Video Showcase).
| Chapter / Era | Commits | Hardware & Core Subsystem | Diagnostic Bottleneck / Dead-End | Engineered Solution |
|---|---|---|---|---|
| 00 / Browser WebGL Prototype | 98e9116 | Client-side Three.js / WebGL (src/main.js, src/carMesh.js) | Browser WebGL lacked hardware ray tracing, multi-million-triangle botanical budgets, and high-res LiDAR shading; cars were blocky procedural extrusions on a blurry vertex-colored heightmap. | Built the initial 5-source telemetry/LiDAR fusion & interactive telemetry HUD in WebGL, then pivoted to dedicated cloud GPU rendering when vmCopilotV710 became available. |
| 01 / Origin & Cloud RT | 98e9116 -> 609a8f1 | AMD Radeon Pro V710 (54 RDNA 3 Ray Accelerators) | 3D terrain scans and WRC telemetry are never distributed together; local Hyper-V VM had no hardware GPU. | Fused 5 open datasets (10 Hz WRC telemetry, 30 Hz UT-ADL RTK-GNSS with EPSG:3301 offset discovery, 1 m LiDAR DTM, Riegl .laz clouds, WMS imagery) and injected gc_11_0_3 firmware into initramfs to bring up Vulkan ray queries. |
| 02 / 260k-Strand Hair Grass | 55f525d, c02c88e | OpenGL 4.6 / ModernGL + 6-layer moss shell | Sparse billboard tufts looked artificial; ModernGL's (LINEAR, NEAREST) texture filter snapped grass along the road every 5.5 m. | Built a 260,000-strand (1.82M-vertex) Cubic-Bezier hair-ribbon groom in a rolling 92 m window with Kajiya-Kay + Marschner shading, 54,000 3D gravel pebbles, and explicit 32-bit float texelFetch spline interpolation. |
| 03 / Rally Estonia Grounding & Card-Tree Pivot | ab66786, bbb12a9 | Native 1920x1080 streaming + shadow FBO fix | glDisable(GL_DEPTH_TEST) prevented shadow_fbo depth writes after frame 0; steering angle was multiplied by 0.075 twice; procedural tree cards looked flat. | Restored shadow-map depth testing and true Ackermann wheel angles, calibrated against 8 Rally Estonia photos, and pivoted away from procedural tree card atlases to 3D scanned botanical meshes. |
| 04 / 4.10M-Tri Botanical Bundle | 2ff3a0d | compile_ultra_highpoly_forest.py (8 species x 3 LODs) | Raw Poly Haven 3D tree GLTFs rendered as dead brown skeletons ('bare brown sticks'). | Solved 4 compounding causes: topological connected-component cluster scaling, morphological solid-color dilation (alpha=255 on 3D needles), Estonian summer chlorophyll re-grading (G/R = 1.30), and lower canopy skirts. |
| 05 / Golden Hour, 682k Yaris & Blackwell | a2aacbf -> 3beb128 | V710 -> A100 -> NVIDIA RTX PRO 6000 Blackwell | Fixed world-space sun pointed from behind the camera on north-west stages; 231k car mesh had faceted arches; V710 hit Hyper-V SR-IOV VF timeout (-110). | Switched to corridor-relative cross-lighting (54 deg azimuth, 13.6 deg elevation) with 3D ray-locked shafts, subdivided the GR Yaris Rally1 to 682,182 curved PN-triangles with native 8K/4K atlases, and migrated to RTX PRO 6000. |
| 06 / 15-Still Forensic Audit (RCA-1..8) | c665256 -> a0b2a0a | 6 parallel engineering threads | 120 px carEllip UV smear; 50% f_tris[::2] rival car decimation; barns buried to their eaves in 1.8x terrain; crystalline pebbles; low banded clouds. | Implemented 12-pass 3D wheel shutter blur, Kevlar wheel liners, UE5 Landmass-style graded building pads, spring-damper camera prop avoidance, two-phase dust + Weber V-curtain water spray, and crown-first 3D Perlin-Worley cumulus clouds at 1.45 km LCL. |
| 07 / Broadcast Director & Shadow-Floor Calibration | 89c9388 -> 63cbb26 | 6-shot camera director + exclusive stream sessions | Full 3D tree shadow maps (FOREST_SHADOW_VERT_SHADER) stacked with Option C's 0.03 ambient shadow floor, crushing forest interiors to ink-black. | Re-engineered terrain/foliage/road shadow floors with open Baltic blue sky-fill, warm bounce, and self-shadow lift while preserving Option C golden-hour contrast and exclusive stream-session ownership. |
| 08 / Survey-Grounded Landscapes & Temporal Stability | edb4b04 | Maa-amet 1 m + ETAK + SYKE 2 m + Metsakeskus + C2 camera | All four stages still looked like the same uniform green tunnel; 345 m vertex teleportation to vec4(2,2,2,1) caused sky gashes and paused-frame flicker (max diff 105). | Built stage-specific national geodata fields, 3-tier Estonian metsaserv edges, frustum-first LOD selection, anticipatory C2 crown-clearance cameras, and proper GPU frustum clipping (paused max diff 105 -> 1). |
| 09 / Telemetry-Coupled 80 Hz Vehicle Physics & Ballistic Suspension | dd02879 | Cloud EGL + Browser WebGL (v710_renderer.py, a100_renderer.py, src/main.js, src/carMesh.js) | Cars felt weightless ('on a turntable'): 2 Hz synthetic straight-line yaw sine wiggle (rut_yaw_kicks), uncoupled lateral crabbing, ignored accx/accy G-sensors, rigid wheel-to-body tilt, and clamped linear spring heave over jumps. | Fused auto-calibrated 10 Hz onboard G-sensors (accx, accy) into an 80 Hz sub-stepped nonlinear single-track bicycle ODE with non-holonomic velocity heading, +0.78 m front-axle drift pivot, parabolic crest flight (-7.2 cm droop, 2.35x Reiger damping), and decoupled 4-corner unsprung wheel knuckles. |
“we have this VM, can we do remote rendering from there - vmCopilotV710”
Before any cloud GPU servers, scanned botanical trees, or hair-ribbon grass existed, the very first version of
TGR-WRT-demo ran entirely in the browser as a client-side Three.js / WebGL application
(src/main.js and src/carMesh.js). Even at this earliest stage, the core analytical idea was already
in place: synchronizing #69 Kalle Rovanperä, #33 Elfyn Evans, and the #00 A100 CUDA Optimal
friction-circle ghost across SS22 Ruuhimäki, SS13 Neeruti, SS14 Elva and SS12 Otepää
with live speed/throttle/brake telemetry traces and a 1 m LiDAR elevation profile scrubber.
Visually, however, the browser WebGL scene was purely schematic: the cars were blocky procedural extrusions with
glowing orange/red proximity rings, the stage road was a flat pink-and-mint ribbon, LiDAR returns were floating colored
dots, and the terrain was a low-resolution vertex-colored green heightmap with zero trees or grass. The turning point came
mid-conversation (left panel below) when we inspected a 500k-triangle Toyota GR Yaris mesh and decided to move rendering off
the browser onto a dedicated cloud GPU VM (vmCopilotV710).
The project began from a fundamental data gap: high-resolution 3D terrain scans and competitive WRC rally
telemetry are never published together. To build a physically grounded 3D Digital Twin for Toyota GAZOO Racing
(TGR-WRT) without proprietary stage packs, we engineered an automated data-fusion pipeline combining
five open scientific and government sources:
#69 Rovanperä, #33 Evans, #00 Optimal):
Synchronized speed, RPM, gear, throttle, brake, steering and 3D GPS traces across SS2/SS5 Elva,
SS3/SS6 Neeruti, SS4/SS7 Otepää and Ruuhimäki, paired with an offline GPU forward-backward
friction-circle optimal-control solver (#00).summer2021-positions.parquet after discovering the exact local-to-L-EST97 coordinate
offset (E = 650000 + x, N = 6465000 + y in EPSG:3301).Because the local workstation environment exposed only a virtual Microsoft Hyper-V Video adapter, we
provisioned cloud GPU instances (vmCopilotA100 and vmCopilotV710). On the AMD Radeon Pro V710,
the Linux kernel initially failed to initialize the GPU with missing firmware amdgpu/gc_11_0_3_mes1.bin (error -2).
Injecting the gc_11_0_3 firmware suite into initramfs unlocked all 54 RDNA 3 hardware Ray Accelerators
(VK_KHR_acceleration_structure + VK_KHR_ray_query), powering a custom C++/Vulkan ray-query bridge
(libv710_vk_rt.so) over a 36,000-triangle terrain/car BLAS. Visually, however, the first live frames were
purely geometric: trees were procedural green cones and the road was a simple ribbon.
“I don't like the grass, this is just tufts here and there. We should have something like hair systems… Which rendering engine are we using actually? Should we switch to unreal or unity?”
Rather than scattering isolated cross-quad grass tufts (`v710_chase_moss_v1.jpg`), we architected a UE5-Groom / Ghost of Tsushima style instanced 3D curved Cubic-Bezier hair-ribbon grass system (while also generating a companion UE5.4 C++/Python project scaffold with Niagara dust and Chaos Vehicle configs):
1.82M vertices): Continuously recycled inside a
camera-locked 92 m rolling field around the car, with species zones (tall golden-headed Calamagrostis reed grass on
outer banks, lush Festuca meadow fescue on ditch slopes, and short Poa pratensis verge turf near the gravel edge).5.5 m. Root cause: ModernGL mapped a 2-tuple filter on a 32-bit float road-spline texture to
(LINEAR, NEAREST), disabling hardware linear interpolation! Switching to explicit sub-texel texelFetch
linear interpolation in GLSL eliminated the stepping completely. We also added 54,000 instanced 3D gravel pebbles
in a 28 m rolling window.
“ok stop, this is a dead end. do extensive research first to see best practices in generating procedural trees, failing that, just download very high poly tree models that fit… abandon these generated trees, they look terrible.”
With the grass carpet working, we audited the renderer against eight real WRC Rally Estonia photographs from
Otepää, Elva and Tartu, upgraded the live stream pipeline from 720p to native 1920×1080 (bbb12a9),
and uncovered two critical rendering bugs:
glDisable(GL_DEPTH_TEST) at the end of
render_frame left depth testing disabled when the next frame's shadow_fbo pass began. Because
OpenGL suppresses depth-buffer writes whenever GL_DEPTH_TEST is disabled, the 2048×2048 shadow map stayed
frozen at 1.0 after frame 0, leaving only a rectangular analytic traceCarBox proxy under the car!0.075 × 0.075): Steering wheel angle was multiplied
by 0.075 both during telemetry ingestion and inside render_frame, reducing a 25° wheel lock to an
imperceptible 0.3°.baltic_3d_tree_atlas_v6) with normal-mapped bark and foliage cards (`v710_v6_low_orbit_trees.jpg`). Under close
orbit inspection, intersecting alpha cards still read as flat paper cutouts. Following the user's pivot, we abandoned
procedural card trees entirely in favor of real 3D botanical geometry.
When we first imported raw 3D tree GLTF models from Poly Haven (`v710_ph3d_chase.jpg` and `v710_ph3d_forest_orbit.jpg`), the result was a visual shock: the trees rendered as dead, skeletal brown sticks. A four-part forensic diagnosis revealed why naive 3D plant imports fail at rally viewing distances:
4.14M individual 1 cm × 2 mm 3D needles (148 m² total needle area). Subsampling to
100k triangles without scaling up the surviving needle clusters reduced canopy area to 3.6 m²![0,0,0,0] pixels) was bound to already-3D-modeled individual needles and leaves. Worse, trilinear mipmapping
averaged the transparent [0,0,0,0] background into the needle RGB texels (`mean_rgb = [27, 27, 15]`), turning
surviving foliage pitch-black at distance.R > G,
[101, 98, 46]) instead of lush July Estonian chlorophyll (G/R = 1.30).y = 0.45, leaving an
empty trunk hallway at the 2.8 m chase-camera eye height.We built scripts/compile_ultra_highpoly_forest.py to compile 13 Poly Haven botanical scans into an
8-species × 3-LOD runtime bundle (4.10M LOD0 triangles, 2,068,548 vertices):
finding exact topological needle/leaf clusters via scipy.sparse.csgraph.connected_components and scaling each
cluster around its 3D centroid, forcing alpha = 255 with 5-pass morphological solid-color dilation, re-grading
foliage to Estonian summer chlorophyll, and grafting lower canopy skirts (`uhp` v1 → v3 → v7).
Next we tackled atmospheric lighting, hero vehicle fidelity, and cloud GPU resilience:
a2aacbf): A fixed world-space sun vector
([-0.64, 0.25, -0.73]) pointed from behind the camera on north-west stages (dot(fwd, sun_dir) = -0.35),
eliminating Henyey-Greenstein forward scattering in dust and foliage translucency. Switching to a stage-corridor-relative
low sun (54° cross-track azimuth, 13.6° elevation, 3800K amber key against 7500K
zenith sky fill) and 3D ray-locked canopy shafts (projecting world points along the sun vector to canopy height
14.5 m) restored dramatic golden side-rim lighting across all four stages.5b03395): We replaced the 231k-triangle car
mesh and downscaled 4K diffuse atlas with crack-free edge-welded Curved PN-Triangle subdivision (682,182 triangles)
and a native 1:1 8192×8192 Albedo + 4096×4096 Tangent-Space Normal + 4096×4096 Packed ORM
material pipeline with dual-lobe carbon-weave clearcoat and micro-flake metallic paint.5276d92, 3beb128): When
vmCopilotV710 suffered a hypervisor SR-IOV VF PCIe reset timeout (error -110), we built the dual-backend
a100_renderer.py / v710_renderer.py parity architecture and migrated onto NVIDIA RTX PRO 6000 Blackwell
with hardware 10-bit AV1 NVENC and WebCodecs low-latency streaming.
Instead of judging quality from a single flattering chase angle, we captured a 15-still forensic matrix covering pack races, close-up wheel wells, rural buildings, roadside props, dry gravel roost, puddle splashes and sky horizons, then dispatched six parallel engineering workstreams to eliminate every artifact:
69d7330, 13cfd20):
Removed the hardcoded screen-space UV ellipse (carEllip) that smeared close-up and pack cars by 120 px; removed
f_tris[::2] 50% triangle deletion on rival cars #33 and #00; corrected tire ground-contact heights
and added radial contact-patch sidewall bulge; enclosed hollow wheel arches with 3D Kevlar liners; and implemented 12-pass 3D
rotational shutter accumulation for spinning wheels.054534f, b617736):
Because terrain relief used a 1.8× vertical exaggeration, a 14 m wide barn placed on a 12° hillside had a
5.3 m elevation delta across its footprint — burying the uphill wall to its eaves (`topic2_08_barn_elva_t77.20.jpg`).
We implemented UE5 Landmass-style smoothstep foundation pad grading in _build_stage plus scanned rural buildings (`B1..B6`).526c405):
Added spring-damper camera prop collision repulsion, replaced orange cylinders with white LLDPE-wrapped silage bales and birch
marker stakes, added sagged 3D catenary stage banners, and flattened jagged crystalline gravel into oblate river pebbles.6a307e2, e741743, 9015b73):
Separated ballistic gravel roost from translucent Beer-Lambert silt wakes, built a dedicated hydrodynamic V-curtain liquid sheet
shader for puddle splashes, and replaced low horizon-hugging cloud bands with a crown-first (hFrac = 1.0 - s) 3D
Perlin-Worley cumulus layer at the 1.45 km Estonian summer lifting condensation level.
As the live viewer matured with a 6-shot broadcast camera director, pursuit-drone aerodynamics
(7fa3d50, 641e666), 10 Hz mistake-zone telemetry overlays (079ac4e), and single-owner stream
session protection (09c4dfa), a subtle lighting interaction appeared when we restored the dramatic
Option C Cinema Golden-Hour Contrast preset (567ea12):
shadow_fbo. Later, full 3D tree shadow casting (FOREST_SHADOW_VERT_SHADER) was
enabled across the 170 m ortho shadow cascades. Stacking real 3D tree shadow maps with Option C's 0.03 terrain shadow
floor, 0.20 AO floor, and 1.15 ACES power curve multiplied shadowed terrain down to 0.6%
albedo before ACES crushed it to pure black!63cbb26): We decoupled shadowed sky-dome diffuse irradiance
from direct sun occlusion — lifting the terrain shadow floor to 0.22, adding Baltic blue sky-fill
(vec3(0.055, 0.068, 0.082)) and warm ground bounce in shadowed foliage, and relaxing ACES shadow toe compression
(`06_fixed_open_shadows_chase_t12.jpg`).
Even with 4.10M-triangle scanned trees, 260k-strand grass, a 682k-triangle car and golden-hour lighting, all four rally stages
still looked like the same uniform green tunnel. Our final investigation — documented in full detail across Sections 03 through 10
of this report — traced that homogeneity to six structural root causes in _build_forest_mesh and the vertex shaders:
_build_forest_mesh ignored LiDAR canopy heights and orthophotos, driving all
four stages with the same 1D trigonometric wave _eval_biome_openness(s_m, side).prob = base_prob * (1 - openness * 0.85)) turned closed forests into sparse golf-course savannas instead of alternating
between 100% open agricultural meadows and tight, closed-canopy stands.w_s Crown-Width Illusion: Height-normalized conifer meshes have a model-space radius of only
0.13..0.25 units, making their true world-space crowns 4× narrower than nominal w_s assumed.0–10 m with a continuous curtain of ditch ferns/shrubs (skirt), grey alder/hazel/willow saplings (mantle), and
interlocking pine/spruce/birch crowns (overstory).[:n2] before frustum culling let
off-screen trees consume half the visible quota, while fol_all[::2] / fol_all[::4] thinned distant crowns.gl_Position = vec4(2.0, 2.0, 2.0, 1.0) stretched boundary triangles across the screen and caused paused-frame flicker
(max diff 105 → 1 after restoring hardware frustum clipping and 1.0 far sky depth).The rest of this report presents the complete geospatial data pipeline, engineering implementation, 54 matched still pairs, and 26 five-second 1080p/30 FPS motion clips proving Chapter 08 in detail.
“the cars move in a way that doesn’t quite seem believable, they don’t seem to have the ‘weight’ behind it. Especially the heading, it seems the cars just turn on a dime, as if the steering wheel directly changes the heading (as if it’s on a platform).”
Inspecting the car’s motion in playback exposed five mathematical causes behind that “weightless / turntable” feel:
(1) lateral offset (lat_offset) and yaw slip (slip_angle) were solved as two uncoupled filters while heading
referenced the static road centerline tangent rather than the car’s displaced velocity vector (“crabbing on a dolly”);
(2) corner-entry yaw added up to 23° of algebraic curvature slip around the center (0,0,0) while
rut_yaw_kicks injected a synthetic 2.0 Hz straight-line yaw sine wiggle (±6.5° to
±22.9° on straightaways); (3) real 10 Hz onboard accelerometer channels (accx, accy)
in the WRC telemetry JSON were ignored (and had inverted axis polarity on ruuhimaki); (4) all four wheels were multiplied
by the sprung chassis_mat (w_world = chassis_mat @ w_local), tilting the tires rigidly off the road when the
body rolled or dived; and (5) crest jumps used a clamped symmetric linear spring (ζ = 0.28) with no parabolic
free-fall or rebound damping control.
In commit dd02879 (scripts/v710_renderer.py, scripts/a100_renderer.py, src/main.js,
src/carMesh.js), we replaced the uncoupled filters with a coupled 80 Hz sub-stepped nonlinear 6-DOF rally physics
engine. See Section 02c below for the full mathematical architecture and 8 synchronized
1920×1080 / 30 FPS A/B video comparisons.
Landscape fidelity only works if the hero Toyota GR Yaris Rally1 (#69) and its pack rivals (#33,
#00) hold up under close-up broadcast scrutiny. Below we zoom into the three most important vehicle-level
engineering breakthroughs — pairing calibrated 1:1 high-magnification macro crops with the full
1920×1080 before/after frames. Click any macro card or close-up to inspect it at up to 500% magnification
in the full-screen lightbox.
| Feature Subsystem | Before (Diagnostic Failure Mode) | After (Engineered 3D Solution) | Commits / RCAs |
|---|---|---|---|
| 01 / Wheel Blur & Wheel-Well Enclosure | 120 px screen-space UV ellipse (carEllip) smeared car bodywork horizontally while 3D wheel geometry rendered in 1 static pass (frozen spokes at 142 km/h) inside hollow open wheel arches. | Removed screen-space carEllip UV smear; added 12-pass 3D rotational shutter accumulation (N_WHEEL_BLUR = 12 over 180 deg shutter angle), radial contact-patch sidewall bulge, and 3D Kevlar inner fender liners. | 69d7330, 13cfd20 (RCA-1, RCA-3, RCA-4) |
| 02 / Mesh Subdivision, Native 8K PBR & Rival Integrity | 231,098-tri base mesh had polygonal wheel arches and a downscaled 4096x4096 atlas; rival cars (#33, #00) used f_tris[::2] 50% triangle deletion, turning bodywork into 'Swiss-cheese' holes. | Edge-welded Curved PN-Triangle subdivision to 682,182 triangles + native 8192x8192 sRGB Albedo + 4096x4096 Normal & ORM atlases with dual-lobe clearcoat; restored full 682k watertight mesh on all rival cars. | 5b03395, 69d7330 (RCA-2) |
| 03 / 5-Part Articulated Wheels, Steering & Braking Dive | Monolithic static GLB mesh could not spin or steer wheels independently; steering angle was multiplied by 0.075 twice (0.3 deg max lock); chassis lacked braking pitch and arch compression. | Connected-component 5-part topological split (body + FL/FR/RL/RR wheels) keeping struts on chassis; true +/-25 deg Ackermann kingpin yaw + cumulative axle spin integral(v/r dt) + telemetry braking pitch/dive. | ab66786, 69d7330, 13cfd20 |
In earlier builds, high-speed motion blur was approximated in the post-process shader using a hardcoded
120 px screen-space UV ellipse (carEllip, RCA-1), while the 3D wheel geometry itself was
drawn in a single static pass (0 rotational shutter samples). At 142 km/h on SS2/SS5 Elva,
this produced two simultaneous visual contradictions:
Panasonic rear wing, #69 door number, and TOYOTA tailgate horizontally across 120 pixels,
yet the white 12-spoke OZ Racing gravel rims remained completely frozen as if parked.RCA-4) & Rigid Tires (RCA-3): Because the raw car
shell lacked inner fender wells, bright sky and background grass leaked straight through the gap behind the front tires, and
the tires floated rigidly without radial load deflection at the gravel contact patch.The Fix: We removed carEllip screen-space smearing from the car body and implemented true
12-pass 3D rotational shutter accumulation (N_WHEEL_BLUR = 12 sub-frame rotor steps distributed
across a 180° shutter angle Δθ = ω · Δtshutter around each wheel’s
3D hub axis), added radial contact-patch sidewall bulge (RCA-3), and sealed all four wheel wells with 3D matte
Kevlar inner fender liners (RCA-4).
Close-up cameras exposed two separate geometric bottlenecks — one on the hero car and a catastrophic decimation bug on the pack rival cars:
5b03395): The original 231,098-triangle
GLB exhibited polygonal, faceted silhouettes along the flared carbon wheel arches and front splitter, while its diffuse
texture had been downscaled to 4096×4096, softening thin sponsor scripts (DMG MORI, CCI,
DENSO, Pirelli). We implemented crack-free edge-welded Curved PN-Triangle subdivision to
682,182 triangles and bound the native 1:1 8192×8192 sRGB Albedo atlas alongside
4096×4096 Tangent-Space Normal and 4096×4096 Packed ORM (Occlusion / Roughness / Metallic)
maps with dual-lobe carbon-weave clearcoat and micro-flake metallic paint.RCA-2, 69d7330):
In pack-chase views, rival cars #33 Evans and #00 A100 Optimal had been rendered with
f_tris[::2] — slicing out every second triangle (50% polygon deletion) as a crude optimization. That left both
rival cars looking like shattered polygon soup with missing doors, roof holes, and broken bumpers (`car_pack_decimation_before.jpg`).
Restoring the full watertight 682,182-triangle mesh on all three cars eliminated every gap.
The raw Toyota GR Yaris Rally1 asset was delivered as a single monolithic static mesh — the wheels, tires, brake
calipers, coilover struts, and chassis were all fused into one vertex buffer. To make the car respond physically to 10 Hz / 30 Hz
WRC telemetry, we engineered a three-stage mechanical articulation pipeline:
scipy.sparse.csgraph.connected_components
combined with axle spatial bounding volumes (z_front = +1.23 m, z_rear = -1.32 m,
y_axle = 0.33 m, r_wheel = 0.35 m), we partitioned the mesh into five independent articulated
groups (body + FL, FR, RL, RR wheel/tire/rim assemblies)
while keeping the upper orange/gold coilover suspension struts rigidly attached to the body shell so they never spin with the
wheel (`yaris_5part_topological_split.jpg`).ab66786): Early builds multiplied
steering wheel telemetry by 0.075 during ingestion and again by 0.075 inside
render_frame (0.075 × 0.075 = 0.005625), clamping a 25° cornering lock down to an invisible
0.3°! Fixing the double attenuation restored full ±25° kingpin yaw articulation
(Y axis) composed with cumulative angular wheel integration θ(t) = ∫(v / rtire) dt around each
hub’s local X axis (`zoom_articulated_wheels_macro.jpg`).69d7330, 13cfd20):
Under heavy braking entering the chicane, longitudinal deceleration drives forward chassis pitch/dive, compresses the front wheel
arch down over the tire, heats the ventilated steel brake discs to a visible 880 °C orange-red thermal glow behind
the calipers, and composes kingpin steering lock with 12-pass 3D rotational spoke blur (`zoom_braking_dive_macro.jpg`).
Even with survey-grounded forests and a 682,182-triangle car mesh, motion playback is only convincing if the car carries
authentic rotational inertia, weight transfer, and unsprung wheel compliance. When we audited why the car previously felt
“weightless / on a platform / turning on a dime,” we traced the artifact to five mathematical shortcuts in the
40 Hz spline pre-integrator and wheel transform hierarchy — and replaced them in commit dd02879 across both the
cloud EGL renderers (scripts/v710_renderer.py, scripts/a100_renderer.py) and the browser WebGL client
(src/main.js, src/carMesh.js).
| Physics Subsystem | Before (40 Hz Uncoupled Linear Filters) | After (80 Hz Coupled 6-DOF Dynamics, Commit dd02879) | Implementation / Evidence |
|---|---|---|---|
| 1 / Onboard G-Sensor Fusion & Auto-Polarity | Ignored 10 Hz onboard accelerometer channels (accx, accy); relied solely on Gaussian-smoothed 10 Hz GPS spline curvature kappa * v_x^2; Ruuhimaki logger had inverted accx polarity. | Loads 10 Hz accx/accy, auto-calibrates axis polarity via dot-product against GPS curvature & pedal effort (fixing Ruuhimaki), and fuses lat_accel = 0.64*(kappa*v_x^2) + 0.36*(accx*g) and lon_accel = 0.45*dv/dt + 0.30*(accy*g) + 0.25*pedals. | StageAssets.__init__, _build_vehicle_physics_splines |
| 2 / Straight-Line Momentum vs. 2 Hz Sine Wiggle | Injected synthetic spatial sine waves (rut_weave + rut_yaw_kicks) that oscillated heading up to +/-22.9 deg at 2.0 Hz on 160 km/h straightaways with zero lateral force. | Removed synthetic straight-line yaw/lateral sine waves; heavy-mass racing-line ODE (omega_n = 2.25 rad/s, zeta = 0.66) tracks straightaways with 1,230 kg momentum. | AB01 (Otepaa t=49-54s), AB07 (Neeruti t=142.5-147.5s) |
| 3 / Non-Holonomic Velocity Heading & +0.78 m Front-Axle Pivot | Uncoupled lateral filter translated the car sideways while pointing parallel to the centerline ('dolly crabbing') and rotated slip around geometric center (0,0,0). | Couples lateral velocity into travel heading (gamma_vel = gamma_road - atan2(lat_vel, v_x)) and shifts drift pivot +0.78 m forward (lat_offset = lat_line + 0.78*sin(beta)) so front tires stay pinned to the inside apex while the rear tail swings wide. | AB02 (Ruuhimaki t=233-238s), AB08 (Otepaa t=327.5-332.5s) |
| 4 / Coupled 80 Hz Single-Track Bicycle Yaw Inertia | Memoryless algebraic drift target (-0.36 * tanh(lat_accel / 5.4)) double-counted turn rate at corner entry ('turning on a dime') and snapped to 0 deg mid-S-bend. | 80 Hz sub-stepped single-track yaw-sideslip ODE with longitudinal weight transfer (lon_wt_ratio), trail-braking turn-in moment (dkappa/dt * brk), and mid-air angular momentum conservation (wn_beta scaled by w_gnd). | AB06 (Elva pack t=22.5-27.5s), AB07 (Neeruti t=142.5-147.5s) |
| 5 / True Ballistic Parabolic Crest Flight & 2.35x Reiger Damping | Clamped symmetric linear spring ([-6.2, +9.5] cm, zeta = 0.28) never entered free-fall (Fz -> 0), clamped droop to -5.6 cm, and pogo-bounced after jumps. | Detects airborne unloading (w_gnd -> 0 over crests), integrates parabolic free-fall with full -7.2 cm Reiger wheel droop, progressive hydraulic bump stops (+4.4 cm), and 2.35x asymmetric rebound-to-bump damping (c_rebound = 23.5 vs c_bump = 10.2). | AB03 (Ruuhimaki jump t=255-260s), AB04 (Elva 177 km/h crest t=106.5-111.5s) |
| 6 / Decoupled 4-Corner Unsprung Wheel Knuckles | Multiplied all 4 wheels by sprung chassis_mat (w_world = chassis_mat @ w_local) with only 55% jounce transfer — tilting wheels rigidly off the ground when the body rolled or dived. | Counter-rotates wheel knuckles against sprung body roll (-0.82 * susp_roll) and pitch (-0.85 * lon_pitch_rad) with MacPherson camber gain (-2.2 deg + 0.52*j_w) and 100% kinematic 4-corner jounce so tires stay locked flat in the 3D gravel ruts. | AB05 (Neeruti 93% brake t=268.5-273.5s), V710CloudRenderer._draw_car |
On fast straightaways and gentle kinks, the previous solver injected spatial sine waves (rut_yaw_kicks and rut_weave) that swung the car’s sideslip between -23.0° and +22.9° (`max |Δβ| = 26.6°`) as if rotating on a motorized turntable. The new 80 Hz coupled engine removes synthetic straight-line sine forcing and drives the heavy-mass racing line (ωn = 2.25 rad/s, ζ = 0.66) strictly from calibrated onboard lateral G (accx) and road curvature.
On ruuhimaki, the raw onboard accx channel had inverted logger polarity and was previously ignored. In the Before clip, the car translates laterally across the road while pointing parallel to the centerline (“crabbing on a dolly”) and pivots around (0,0,0) while bouncing on fake vertical bump waves (+8.1 cm). In the After clip, auto-calibrated accx (1.01G) fuses with curvature, travel heading aligns with the non-holonomic CoM velocity vector (γvel = γroad − atan2(˙ey, vx)), and the +22.5° power slide pivots around the biting front axle (+0.78 m ahead of CoM) so the front wheels stay pinned on the inside apex groove while the rear tail swings wide.
Over the signature Finnish Ruuhimäki jump crest at 162 km/h, the previous model used a clamped symmetric linear spring (ζ = 0.28, droop limited to −5.6 cm) that never entered free-fall and pogo-bounced after landing. The new 80 Hz ballistic solver detects spring unloading (wgnd → 0), launches the chassis along a true parabolic arc (+9.1 cm heave) with all four wheels extending to −7.2 cm full Reiger droop below the sills, conserves angular momentum in mid-air, and absorbs the multi-G landing impact into progressive hydraulic bump stops (+4.4 cm) with 2.35× rebound-to-bump damping (crebound = 23.5 vs. cbump = 10.2).
Framed from a low 3/4 wheel-arch tracking camera at 177 km/h on SS2 Elva, this pair isolates high-speed vertical unloading followed by compression into a fast left-hander. Notice how the After suspension extends all four tires to full −7.2 cm rebound droop over the crest and then compresses cleanly into the outside-right wheel arch without lifting the inside tires off the gravel bed.
Under 93% trail-braking and −0.81G longitudinal deceleration into the Neeruti corner, the car pitches +3.8° nose-down and leans −5.6° in roll. Previously, w_world = chassis_mat @ w_local tilted all four wheels rigidly with the sprung body using only 55% jounce transfer — tilting the rear tire sideways off the road. In dd02879, the wheel knuckles counter-rotate against sprung body roll (−0.82 · susp_roll) and pitch (−0.85 · lon_pitch) with 100% kinematic 4-corner jounce ([−7.2, +4.4] cm), keeping every tire planted flat in the 3D gravel ruts while the body visibly dives and rolls above them.
Entering the Elva chicane with all three pack cars visible (#69 Rovanperä, #33 Evans, and #00 Optimal), the uncoupled model swung #69 through a 34.5° sideslip discrepancy with abrupt center-pivot yaw snaps. The coupled 80 Hz solver transfers weight forward under braking (lon_wt_ratio), generates a physical trail-braking turn-in yaw moment (dκ/dt · brk), and coordinates independent racing lines and suspension dive across all three cars.
Through the 181 km/h Neeruti S-bend transition, the memoryless algebraic curvature formula in Before collapsed sideslip to −0.9° right at the inflection point, making the car snap flat as if on rails before jerking into the next bend. In After, the coupled single-track bicycle yaw ODE carries rotational inertia (Izz) across the inflection point (+19.5° sideslip, +0.39 m centrifugal exit drift) with asymmetric roll damping (cr = 11.8 load / 17.5 recovery).
The highest-energy combined braking and cornering sequence across all four stages (187 km/h, 91% brake pressure, −1.10G longitudinal deceleration, +0.73G lateral acceleration). Demonstrates simultaneous +3.9° nose-dive, −5.6° outside roll lean, −22.9° front-axle pivot drift, opposite-lock counter-steering, and decoupled 4-corner wheel articulation.
The most consequential correction to the proposal was to stop treating *_ndsm.jpg as a
calibrated metre-scale height raster. A plausible grayscale or coloured height preview does not establish
its scale, no-data convention, legend, acquisition date or accuracy. The implementation downloads and
reads the numeric Estonian GeoTIFFs, and explicitly treats Finland's source as classified height bands.
| Scene component | Grounding | Modeled or uncertain part |
|---|---|---|
| Route and replay | Stored telemetry positions, elevations and replay samples. | A trace is not a survey of road edges. Road width/cross-section, gravel detail and reconstructed suspension/pose are renderer models. |
| Relief | Stored terrain, now a denser numeric Estonian DTM derivative where acquired. | The existing default vertical presentation exaggeration is 1.8x. Interpolation, local road/ditch engineering and mapped foundation/water adjustments are not pristine survey values. |
| Woodland and openings | ETAK polygons, numeric Estonian canopy heights, Finnish classified cover/heights and qualified stand inventory. | Individual stem positions, species outside inventory, crown shape and understory detail remain deterministic procedural proxies. |
| Aerial colour | Existing stage orthophoto/mosaic files, preserved byte-for-byte through the experiment. | The repository has both a Bing VirtualEarth tile downloader and an Estonian Maa-amet WMS builder that can write the same filenames. Comments alone do not prove which run produced a stored JPEG or its vintage. |
| Buildings and water | Mapped footprints and shorelines. | Generic building appearance, water elevation, surface/reflection response and historic contemporaneity are not measured here. |
| Atmosphere and props | Preserved completed lighting, generic licensed scene assets and contextual placement rules. | Golden-hour direction/look, HDRI horizon, weather, local rocks/deadwood/signs and camera flight are illustrative, not recovered rally-day observations. |
Bing Maps: no Bing runtime feed or new Bing acquisition was introduced in this change. The stored aerial mosaics still contribute limited macro colour. New geographic structure comes from the explicit official canopy/terrain/semantic products below, not an assumed Bing classifier or photogrammetric reconstruction.
| Stage | Canopy / semantic source | Height qualification | Coverage definition | Vegetation instances | Rendered mapped features |
|---|---|---|---|---|---|
| Ruuhimäki | 2 m SYKE height classes + stand inventory | Class midpoints; >20 m refined where inventory exists | 100.0% of sampled cells in the 430 m belt | 517,911 | 232 proxies / 1 polygon parts |
| Neeruti | 1 m numeric nDSM + ETAK | 17 route-adjacent tiles | 100.0% of sampled cells in the 430 m belt | 315,013 | 94 proxies / 33 polygon parts |
| Elva | 1 m numeric nDSM + ETAK | 29 route-adjacent tiles | 100.0% of sampled cells in the 430 m belt | 423,672 | 270 proxies / 63 polygon parts |
| Otepää | 1 m numeric nDSM + ETAK | 36 route-adjacent tiles | 100.0% of sampled cells in the 430 m belt | 651,447 | 251 proxies / 103 polygon parts |
| Earlier assumption | Verified interpretation / implemented response |
|---|---|
| No Finnish class-6 returns means no buildings | False. The acquired Finnish semantic product contains building footprints near the route. Absence of a class in a thinned cloud is not an absence certificate. |
| JPEG values directly decode to exact canopy metres | Not established. No deployed metric height uses the proposed grayscale or colour-ramp formula. |
| Removing seven-tile selection automatically proves complete raw coverage | It removes an artificial limit. Coverage still depends on tile availability, clipping, classification and sampling. The new raster belt is validated separately. |
| Vegetation-return shares measure forest area | They do not. Earlier 69.2% / 37.1% / 31.3% / 35.7% figures describe stored return samples, not land-area percentages. |
| Dropping half the components makes a crown exactly half transparent | False. Projected coverage, overlap, alpha filtering, viewpoint and lighting matter. The selected distant representation avoids that extra component removal. |
| Every forest edge must be an opaque lower-ten-metre wall | Too categorical. The implementation preserves mapped openings and uses layered, locally varied edge growth rather than enforcing a universal wall. |
The north–south registration issue was genuine: raw image row zero and negative world Z both
denote north, but the terrain's V coordinate was inverted. Terrain, landscape sampling and grass now
share z / worldDepth + 0.5. This fixes the convention; it does not turn an approximately
fitted aerial mosaic into an independently surveyed orthorectification.
Data coverage above means finite numeric/classified samples within the specified belt. It is not a survey-accuracy score, a count of individual crowns, or a directly comparable denominator to the old point-cloud station-coverage audit.
Numeric terrain and canopy; semantic vectors; Finnish height classes and inventory; existing route and imagery.
Stand footprint, height prior, validity, metric edge distance, land cover, moisture/tint and provenance.
Overstory, edge mantle, regeneration, ground/grass, mapped foundations and water, contextual road furniture.
Spatial indexing, conservative bounds, stable LODs, cutout filtering, shared wind, planned cameras and measured output.
Overstory candidates use a 3.8 m world-cell lattice with deterministic jitter and bounded density priors. A separate 2.8 m edge layer supplies shrubs, saplings and small broadleaves; a 7.5 m interior layer introduces sparse regeneration. These spacings are implementation choices, not claims that every Estonian or Finnish forest has one universal stem density. Composite sapling and shrub assets also mean an instance count is not a stem count.
Canopy height and stand boundaries determine eligibility. Water, buildings, rock, roads and the engineered road clearance exclude unsuitable placements. Source-supported isolated patches can survive in open land. The model uses a jittered area process, not a fully calibrated Thomas-process, crown-competition or silvicultural growth simulation.
Meshes are normalized by height, so horizontal instance scale is not a crown diameter. The renderer measures the loaded geometry and derives scale from a desired crown span and actual model radius. The table records the 95th-percentile foliage-vertex radius and maximum radial extent. It is not a leaf-area-weighted botanical crown measurement.
| Species / asset slot | Radius p95 (model units) | Maximum radius | Normalized height |
|---|---|---|---|
| 0 | 0.205 | 0.263 | 1.000 |
| 1 | 0.173 | 0.226 | 1.000 |
| 2 | 0.174 | 0.236 | 1.000 |
| 3 | 0.368 | 0.581 | 1.000 |
| 4 | 0.377 | 0.499 | 1.000 |
| 5 | 0.634 | 0.861 | 1.000 |
| 6 | 0.242 | 0.385 | 1.000 |
| 7 | 0.572 | 0.916 | 1.000 |
Per-species spatial indexes restrict the query region, then conservative union bounds are frustum-culled before LOD selection. There is no nearest-N species quota silently reshuffling visible trees. Projected-size selection uses hysteresis; discontinuous seeks/cuts reset it, while a stationary pause does not force a new geometry choice. Complete authored LOD2 foliage is retained at distance, avoiding the previous derived half/quarter-component thinning.
Coverage factors are measured per texture layer and mip level, with bounded scale and an anisotropy-aware lookup. They support alpha-to-coverage and 8x MSAA; they are not an exact optical-depth invariant. Colour, shadow and velocity passes share deformation and cutout logic. Fine ground noise is attenuated when undersampled, and foliage chromatic variation no longer slides through a high-frequency world-space sine as the leaves move.
Detailed crowns fade over 340–430 m into mapped ground treatment and haze. An elevated, source-shaped canopy-sheet prototype was tested and rejected: it produced faceted, blanket-like horizons in wide views. It is not shipped or counted as a visual improvement. Proper distant volumetric/impostor crowns remain future asset work.
A separate wide-view defect came from an existing vertex-culling shortcut: connected terrain and road vertices beyond 345 m were moved individually to one clip-space point without initializing their varyings; the scenery shader had the same defect at 180 m. Mixed triangles stretched across the image. Ordinary GPU primitive clipping now preserves the surface, and the sky fills only untouched far depth. An actual GPU regression checks surface/sky ordering at 3 km.
The broad road-height floor is removed. A local road/ditch correction ends at 16 m, with modest procedural surface displacement and separate mapped foundation/water adjustments. Vertical presentation exaggeration remains explicit. Estonia uses a denser derivative from the numeric DTM; Finland retains the existing terrain source with denser interpolation, not a newly surveyed 1 m Finnish surface.
Terrain and grass read the same woodland/moisture/tint masks. Open agricultural ground, woodland litter and moist edges therefore change together. Mapped buildings receive contained generic proxies, not automatically generated silos, tractors and companion buildings. Water follows mapped polygon boundaries; estimated levels and approximate reflections are qualified below. Signs and anti-cut bales are sparse curvature-derived illustrations, not claimed observations of the historic course.
Road wetness is derived from smoothed vertical sags and canopy shade. The GPU road surface and CPU tire emissions use the same profile. This removes the stage-independent periodic basin field without pretending to know the actual rally-day rainfall or drainage.
Automatic rigs smooth the horizontal reference with acceleration/jerk checks, anticipate conservative crown and static-scenery bounds, and construct a vertically speed-bounded clearance majorant. The resulting C2 path is independent of render wall time and frame partition. A complete-route 120 Hz audit samples terrain, wind-inflated crown cylinders and building/scenery bounds before acceptance. The old independent gimbal is replaced by subject-directed aiming; manual bumper/orbit controls remain.
Conservative full crown cylinders can unnecessarily lift the camera above empty space beneath high branches. That is a deliberate safety/appearance trade-off, visible in the separately labelled native clips. It is not a proof of triangle-level collision freedom or a complete robotics planner.
These are real GPU captures, not generated after-images or retouched illustrations. Main comparisons lock the numeric eye, target, FOV, up vector, grass focus, lighting preset and simulation timestamp. Embedded previews are resized JPEGs; the original lossless PNG links retain full output resolution.
Exact baseline camera and golden lighting. The geographic placement and ground treatment change; the framing does not.
The actual stand footprint changes where trees stop. This is not a blanket density boost.
A mapped open section no longer inherits a synthetic forest wall, fence and repeated standing-water pattern.
The same world-space field controls relief, woodland and understory. Scene identity is not injected by random roadside props.
Source-selected landmark location; exact baseline camera, simulation time and lighting. Building appearance and water level remain modeled.
Source-selected landmark location; exact baseline camera, simulation time and lighting. Building appearance and water level remain modeled.
Source-selected landmark location; exact baseline camera, simulation time and lighting. Building appearance and water level remain modeled.
Source-selected landmark location; exact baseline camera, simulation time and lighting. Building appearance and water level remain modeled.
There is no selection rule that discards an inconvenient view. Some old poses are inappropriate after geographic vegetation is restored; native-camera evidence is shown separately.
Locked baseline view, chase, t=12.35s.
Locked baseline view, heli, t=12.35s.
Locked baseline view, chase, t=35.00s.
Locked baseline view, heli, t=35.00s.
Locked baseline view, chase, t=77.20s.
Locked baseline view, heli, t=77.20s.
Locked baseline view, chase, t=166.66s.
Locked baseline view, heli, t=166.66s.
Locked baseline view, chase, t=273.31s.
Locked baseline view, heli, t=273.31s.
Locked baseline view, chase, t=12.35s.
Locked baseline view, heli, t=12.35s.
Locked baseline view, chase, t=35.00s.
Locked baseline view, heli, t=35.00s.
Locked baseline view, chase, t=77.20s.
Locked baseline view, heli, t=77.20s.
Locked baseline view, chase, t=154.62s.
Locked baseline view, heli, t=154.62s.
Locked baseline view, chase, t=253.57s.
Locked baseline view, heli, t=253.57s.
Locked baseline view, chase, t=12.35s.
Locked baseline view, heli, t=12.35s.
Locked baseline view, chase, t=35.00s.
Locked baseline view, heli, t=35.00s.
Locked baseline view, chase, t=77.20s.
Locked baseline view, heli, t=77.20s.
Locked baseline view, chase, t=190.65s.
Locked baseline view, heli, t=190.65s.
Locked baseline view, chase, t=312.67s.
Locked baseline view, heli, t=312.67s.
Locked baseline view, chase, t=12.35s.
Locked baseline view, heli, t=12.35s.
Locked baseline view, chase, t=35.00s.
Locked baseline view, heli, t=35.00s.
Locked baseline view, chase, t=77.20s.
Locked baseline view, heli, t=77.20s.
Locked baseline view, chase, t=266.82s.
Locked baseline view, heli, t=266.82s.
Locked baseline view, chase, t=437.59s.
Locked baseline view, heli, t=437.59s.
Locked baseline view, chase, t=17.80s.
Locked baseline view, chase, t=22.50s.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Native 3840 x 2160 output from both versions; exact baseline pose. Preview scaled for the page.
Same geographic candidate under other existing lighting presets. These are smoke checks, not a second lighting redesign or proof of every weather/time combination.
Every formal clip has 150 individually rendered output frames at
t_start + n / 30, 1920 × 1080 and constant 30 FPS for five seconds.
Independent local decoding checks frame count, frame rate and duration. No repeated stills, generated
intermediate frames or slowed playback were used to meet that specification.
The report players use size-bounded H264 reading copies, still 1920 × 1080, 150 frames and 30 FPS. Before and after use the same CRF24 / 4 Mbit/s maximum encoding treatment, with no retiming, interpolation, grading or geometry alteration. These compressed previews are not used for subpixel flicker measurements; linked originals and host-retained lossless frames remain the acceptance evidence.
Four additional moving-camera A/Bs reuse every baseline camera pose. The ten changed-path native motion clips instead demonstrate the integrated renderer plus the safer camera. Those are not isolated material comparisons. Their simulation times, seed and output cadence remain matched.
The files contain exact matched camera/time samples. Side-by-side browser playback is only approximately synchronized; it is not the frame-level measurement.
This clip deliberately changes the camera. It demonstrates the integrated scene and planned flight, not an isolated shader improvement.
| Fixed-path control | Original baseline | Same-path candidate |
|---|---|---|
| Ruuhimäki | Baseline clip | Exact-camera candidate clip |
| Neeruti | Baseline clip | Exact-camera candidate clip |
| Elva | Baseline clip | Exact-camera candidate clip |
| Otepää | Baseline clip | Exact-camera candidate clip |
The exact-old-path Neeruti control ends inside restored foliage. That uncorrected legacy pose is retained deliberately to keep the A/B honest; it is not the shipped automatic camera path. The native Neeruti clips use the clearance-planned rig and remain unobstructed in the sampled motion views.
| Exact-camera chase control | Baseline flow p95, median | Candidate flow p95, median |
|---|---|---|
| Ruuhimäki | 26.0 | 24.0 |
| Neeruti | 20.0 | 17.0 |
| Elva | 24.0 | 25.0 |
| Otepää | 28.0 | 24.0 |
| Native / wind case | Camera treatment | Before | After | Median render cost | Flow residual p95 (median) |
|---|---|---|---|---|---|
| ruuhimaki_motion_chase | native camera path | Baseline 1080p | Candidate 1080p | 64.7 ms | 17.0 |
| ruuhimaki_motion_heli | native camera path | Baseline 1080p | Candidate 1080p | 55.1 ms | 13.0 |
| neeruti_motion_chase | native camera path | Baseline 1080p | Candidate 1080p | 70.7 ms | 14.0 |
| neeruti_motion_heli | native camera path | Baseline 1080p | Candidate 1080p | 50.7 ms | 17.0 |
| elva_motion_chase | native camera path | Baseline 1080p | Candidate 1080p | 42.3 ms | 20.0 |
| elva_motion_heli | native camera path | Baseline 1080p | Candidate 1080p | 34.0 ms | 21.0 |
| otepaa_motion_chase | native camera path | Baseline 1080p | Candidate 1080p | 48.7 ms | 19.0 |
| otepaa_motion_heli | native camera path | Baseline 1080p | Candidate 1080p | 38.6 ms | 16.0 |
| neeruti_near_motion | native camera path | Baseline 1080p | Candidate 1080p | 82.5 ms | 14.0 |
| elva_near_motion | native camera path | Baseline 1080p | Candidate 1080p | 51.7 ms | 18.0 |
| elva_wind_stationary | matched baseline poses | Baseline 1080p | Candidate 1080p | 75.4 ms | 7.0 |
The flow figure is a within-clip diagnostic: quarter-resolution luminance is compared after bidirectional-consistency-filtered optical-flow warping, excluding invalid correspondence. Wind, self-shadowing, occlusion and changing scene content remain confounders. It is not a photorealism score, a universal flicker threshold or proof that a changed camera caused a renderer improvement.
373,804 sampled default-rig poses across complete routes produced 0 proxy intrusions. These are conservative wind-inflated crown/scenery bounds, not exact triangles or a certificate covering every future camera control. Inspected motion evidence and the full-route geometric audit serve different purposes.
| Stage | Rig | 120 Hz samples | Proxy intrusions | Max horizontal a (m/s2) | Max vertical v (m/s) | Max vertical a (m/s2) |
|---|---|---|---|---|---|---|
| ruuhimaki | chase | 39,998 | 0 | 14.48 | 6.46 | 3.18 |
| ruuhimaki | heli | 39,998 | 0 | 15.33 | 6.76 | 3.16 |
| neeruti | chase | 37,108 | 0 | 17.03 | 4.52 | 3.10 |
| neeruti | heli | 37,108 | 0 | 13.01 | 6.59 | 3.56 |
| elva | chase | 45,757 | 0 | 16.71 | 5.61 | 3.29 |
| elva | heli | 45,757 | 0 | 14.14 | 6.03 | 3.16 |
| otepaa | chase | 64,039 | 0 | 15.61 | 6.04 | 3.83 |
| otepaa | heli | 64,039 | 0 | 15.99 | 6.01 | 3.59 |
Repeated fixed-time renders over the 42 main candidate stills produced median RGB mean absolute difference 0.000001 on the 0–255 scale; maximum observed channel difference was 3. Camera partition/seek determinism, pause versus cut handling, stable instance identity and frustum/buffer behavior are also tested. A finite corpus cannot establish that all possible views are artifact-free.
The actual Spain allocation is NVIDIA RTX Pro 6000 Blackwell DC-2-48Q/PCI/SSE2. It is a 48 GB vGPU profile, not the 24 GB guest used in the older fidelity research and not an assumed full 96 GB physical card. More VRAM does not imply free geometry, shading or encoding throughput.
| Stage | Baseline same-time chase | After exact old camera | After native planned camera |
|---|---|---|---|
| Ruuhimäki | 36.3 ms | 197.0 ms | 64.7 ms |
| Neeruti | 36.3 ms | 184.6 ms | 70.7 ms |
| Elva | 40.6 ms | 66.8 ms | 42.3 ms |
| Otepää | 35.6 ms | 178.0 ms | 48.7 ms |
Each value is the median direct-render plus blocking BGR-readback wall time over the corresponding warmed 5 s clip, excluding its first frame. Encoding, download, buffering and browser presentation are not included. Camera changes alter overdraw and visible geometry, so the native column is not an isolated engine optimization result. The exact-old-path column is the controlled workload comparison.
The candidate loads all four stages in approximately 61.0 s in this capture process. First-frame camera planning and cold seek work are separately visible in raw timings; they are not advertised as warmed streaming throughput. The production process remained resident during isolated captures, and these are vGPU observations rather than exclusive physical-card benchmarks.
name, memory.total [MiB], memory.used [MiB], driver_version NVIDIA RTX Pro 6000 Blackwell DC-2-48Q, 49152 MiB, 13153 MiB, 595.91.07
The existing viewer keeps native resolution and adapts timestamp sampling with buffering and congestion headroom. It is allowed to deliver fewer than 30 live FPS for these quality-priority settings. Smooth offline clips establish sampled visual behavior, not a network or real-time performance guarantee.
The checked-out viewer's old lease-style requests were incompatible with the finished exclusive-session backend. Its matching registered-client implementation and tests were integrated from the completed rollout rather than weakening the guard. Moving JPEG quality is raised from 56 to 92, matching paused quality; cadence still adapts. This increases delivery cost and does not change the offline render corpus.
| JPEG quality | Frame bytes | RGB RMSE (0-255) | RGB MAE (0-255) |
|---|---|---|---|
| 56 | 228,848 | 6.120 | 4.296 |
| 92 | 633,675 | 4.116 | 2.800 |
Compression error is not a temporal stability or photorealism score. The final live owned-session check and browser playback verify the actual delivery path separately.
The first geographic candidate still put an old aerial eye inside a crown. It was not accepted as a finished improvement, and trees were not deleted to clear the lens. The new path planner addresses that presentation failure.
The second pilot demonstrates clearance at the same route time, but changes the eye and framing. It must not be used as evidence that a shader alone fixed the view.
The first large candidate corpus passed media-format checks but failed visual review. Wide landmark shots exposed a faceted elevated canopy sheet and missing/stretched ground at the old 345 m vertex boundary. That corpus was retained, not promoted. Removing only a fragment cutoff and moving the sky to true far depth was insufficient. Fixing connected-mesh clipping removed the paused-frame instability; a separate on/off ablation then showed that the canopy sheet still looked worse than the underlying mapped terrain, so it was removed rather than disguised.
In the controlled Elva fixed-time probe, the rejected version differed by up to 105 channel levels, with 0.0702% of pixels exceeding an 8-level maximum-channel difference. After the clipping fix, eight repetitions in each of five pass ablations had maximum difference 1/255 and no pixels above that diagnostic threshold. Particle-state hashes and LOD counts stayed fixed in both versions. The final main-corpus repeated-render numbers are reported separately above; this is not a universal motion-flicker score.
| Proposal / research direction | Decision and reason |
|---|---|
| Decode every nDSM JPEG into exact metres | Rejected. Numeric rasters/classified products were acquired instead; preview interpretation was not calibration. |
| Apply universal species percentages and 2.2-3.8m spacing everywhere | Replaced with source footprints, measured model bounds and bounded stand/region priors. Exact species and density outside inventory remain inferred. |
| Guarantee exact optical depth by radially inflating crowns | Not adopted as a guarantee. Complete distant authored geometry, measured cutout correction and empirical motion evidence are more defensible. |
| Push arbitrary woody vertices along normals to reattach floating leaves | Not applied. That is not an attachment-aware repair. Existing finished twig/leaf asset work is preserved; extreme close-ups remain asset-limited. |
| Force an opaque lower-ten-metre wall at every forest edge | Rejected as a universal ecological rule. Edge structure is layered but openings and locally sparse woodland are preserved. |
| Acquire all available raw LiDAR and restricted Finnish dense data immediately | Not necessary for this implementation. Open numeric/semantic derivatives supplied the required belt. Restricted/paid acquisitions and permissions were not assumed. |
| Make 1m terrain everywhere or claim centimetre accuracy | Not claimed. Source resolution, output derivative resolution, rendering grid, interpolation and vertical exaggeration are distinct. |
| Treat mapped buildings as exact historic farmsteads | Rejected. Generic contained proxies replace fabricated compounds; real roof/facade reconstruction and historic inventory remain future work. |
| Infer actual water level/weather from a polygon or a sag | Not claimed. Levels and wetness are models; shoreline geometry is the mapped component. |
| Ship a lower-FPS renderer but hide unstable pacing with offline clips | Rejected. Offline sampling, render cost and live delivery are reported separately; existing adaptive pacing and ownership guards are preserved. |
Surveys and inventories span different years and are not uniformly contemporaneous with the rally replay. Finnish height classes can include power lines or other objects; nDSM includes elevated non-vegetation unless excluded by semantics. The regional species proxies are not a botanical classification of every crown.
The renderer remains a custom OpenGL raster system. Existing Vulkan car-query initialization and shader comments mentioning PCG, Lumen or ray tracing do not establish whole-scene ray-traced forest lighting. Water uses approximate environment/analytical response. Distant individual crowns beyond the detailed range and the finite mapped-extent/horizon transition remain limitations; no rejected canopy sheet is hidden in the final build. No universal photorealism percentage or unmeasured sub-millisecond cost is invented.
The baseline is the integrated, actually served Spain file bundle, not an assumed single Git revision. The completed Option C lighting plus lifted-shadow finish and current stream ownership were preserved. Before capture, source was copied and hashed into an immutable isolated baseline. No production source was overwritten to obtain a favourable experimental frame.
Baseline renderer SHA-256: 291aa27b15bd7e64183f5bcf49baa4c26b79302a458c665148f855ed283a5d96 Candidate renderer SHA-256: 902aa20dc624b6509fab76f010fad33cd5b7b17b2d60f45bef08ffbe46509165
| File / surface | Responsibility |
|---|---|
| scripts/build_landscape_data.py | Numeric/classified acquisition, exact georeferencing, semantic masks, source dates/checksums and validated derivative coverage. |
| scripts/landscape_field.py | Pixel-centre versus endpoint sampling, deterministic area placement, mapped foundation grading, shoreline triangulation and shared sag/shade wetness. |
| scripts/landscape_camera.py | Anticipatory position paths, conservative crown/scenery clearance and full-route 120Hz audits. |
| scripts/v710_renderer.py + scripts/a100_renderer.py | Byte-identical stage/runtime integration; local relief correction; data-driven scene; culling, bounds and complete distant foliage; native camera integration. |
| scripts/foliage_geometry.py | Measured crown dimensions, bounded per-mip cutout coverage factors, and pause/cut temporal state. |
| scripts/photoreal_terrain.py | Shared terrain/grass fields, stable fine shading, consistent foliage passes, valid connected-mesh clipping and mapped water. |
| scripts/rally_surface.glsl + scripts/rally_particles.py | Matching GPU water/moisture and CPU tire emission input. |
| scripts/build_rally_data.py | Removal of the seven-tile raw-cloud cap and physical HAG preservation. This does not claim all raw source clouds were freshly rebuilt here. |
| scripts/v710_server.py + stream_sessions.py + photoreal_post_fx.py | Compatible served runtime; ownership and earlier lighting/motion work preserved. The sky's depth placement is corrected without retuning its lighting. |
| scripts/landscape_evidence.py + collect_landscape_evidence.py | Recorded camera poses, controlled offline frames, independent media verification, source hashes and capture manifests. |
python -m pip install -r scripts\requirements-landscape.txt python scripts\build_landscape_data.py --cache "$env:LOCALAPPDATA\TGR-landscape-cache" python -m unittest discover -s tests -p "test_landscape*.py" python scripts\test_render_integration.py node --test tests\streaming.test.js # On an isolated compatible GPU host: # PYTHONHASHSEED=69 and TGR_STANDALONE_TEST=1 # landscape_evidence.py --source [snapshot scripts] --assets [stage/model assets] # --output [corpus] --label [label] --reference [baseline manifest] # --reference-stills-only: moving native camera clips differ; wind control stays locked.
New geography/camera contracts, the real legacy LiDAR packing function with eight synthetic tiles and unscaled tall returns, renderer parity, particle behavior, stream ownership, shader linkage and actual GPU readback are exercised. The corrected GPU preflight ran 35 tests with two Node-only skips; the Windows streaming suite ran 31 tests. A missing remote HTTP test transport was installed and the preflight rerun, not silently skipped. The compatible latest-viewer registration and quality-92 delivery path are tested separately from the frozen renderer. Diagnostics do not replace the rendered evidence.
Accumulated rejected and superseded lossless experiments filled the host's 29 GB system volume. The encoder failed, so that attempt was not accepted. Spain was recovered, only our superseded experimental lossless files were reclaimed, and baseline/final-source lossless evidence was kept. All four derived inputs and every frozen renderer/shader/helper hash were checked before the formal candidate campaign was restarted. No renderer change was hidden in this operational recovery.
The capture tool now checks a conservative per-clip storage budget plus a 2 GB disk reserve and publishes explicit encoder-failure progress instead of leaving a stale running marker. The host restart changed the temporary tunnel hostname; the compatible viewer's default and recovery manifest were updated together. A live host recovery is not itself evidence of a visual improvement.
{
"status": "not promoted"
}
Promotion is conditional on the accepted evidence. Both runtime and asset copies must contain the same renderer/shader/helper versions; stream-session ownership stays intact. A rollback source snapshot is retained. Live anonymous frame and batch requests must still be rejected, and an owned-session capture must decode at native resolution.
National geospatial products retain their own attribution and terms. Existing imagery remains subject to its separate provider terms; the JPEG alone does not establish provider or acquisition vintage. Poly Haven-derived scene assets retain their existing licensing. No third-party rally reference photograph is republished here as an unverified licensed source.
All 112 original PNG stills and 26 full-resolution original MP4 clips are hosted on this site. The report players use smaller reading copies; lossless FFV1 masters remain privately retained. Browse the original video library or download the public evidence catalogue with checksums. Operational filesystem paths, remote-host access details and private workspace metadata are not published.
{
"status": "verified",
"renderer_sha256": "902aa20dc624b6509fab76f010fad33cd5b7b17b2d60f45bef08ffbe46509165",
"still_count": 112,
"clip_count": 26,
"output_motion_frames": 3900,
"public_originals": "See evidence.json for hosted originals and SHA-256 checksums."
}{
"ruuhimaki": {
"grid_shape": [
3136,
2212
],
"pixel_m": [
1.9997603978300182,
1.9995854591836735
],
"canopy_corridor_coverage": 1.0,
"observed_height_percentiles_m": [
7.5,
12.5,
23.0,
25.875
],
"unknown_height_cells": 0,
"source_feature_candidates": {
"building": 251,
"water": 1
},
"ndsm_tiles": 0,
"crs": 3067,
"sources": [
{
"sha256": "9bd1301e6d23be9bf9e0a17a1ae857ab926e74df4d19e9e24bbb03131f7f81ba",
"file": "ruuhimaki_landcover.tif"
},
{}
]
},
"neeruti": {
"grid_shape": [
1641,
1877
],
"pixel_m": [
1.9990783164624402,
1.9990067032297378
],
"canopy_corridor_coverage": 1.0,
"observed_height_percentiles_m": [
5.31640625,
14.9453125,
23.78125,
28.671875
],
"unknown_height_cells": 227072,
"source_feature_candidates": {
"building": 108,
"water": 33
},
"ndsm_tiles": 17,
"crs": 3301,
"sources": [
{},
{},
{},
{},
{},
{},
{},
{
"sha256": "36ed0c4b3673749133b22a5c9a78b4b5d1638e037989354a7bae80b3391574a9",
"file": "442647_ndsm_1m.tif"
},
{
"sha256": "8900eeac6c0bd97eb349c08c2b443c15a99a99a24e1ff3f2d313cc4da6972ec4",
"file": "442648_ndsm_1m.tif"
},
{
"sha256": "33292bf46199b2a9abac7b1e2bfebbb2fe0963e3b3e07b74ece478d3357fdd46",
"file": "442649_ndsm_1m.tif"
},
{
"sha256": "a7313331b02245f57c703062e148974c1c5567853556e9d037d431e8af04ee63",
"file": "442650_ndsm_1m.tif"
},
{
"sha256": "4b88971fa8b83331dcd335304e3fd28d378110595c79f8814d3f4fecaba50e0b",
"file": "443647_ndsm_1m.tif"
},
{
"sha256": "9ea46ffcb28f4b61418ce06cb8d338962984e49215a26c67dc54835b23cd198b",
"file": "443648_ndsm_1m.tif"
},
{
"sha256": "2a34903d8fe8cc2098785e5184e5ef709d33d97ef8981c57364d0b90c8c9befa",
"file": "443649_ndsm_1m.tif"
},
{
"sha256": "f73ec609213a1333d4a7375cdfcec22d9a53885e67829c65e38d6d88c8c3e943",
"file": "443650_ndsm_1m.tif"
},
{
"sha256": "d0819f652aff937610125faa0f404cc8314d4ac46ada2b957d537be43fab935c",
"file": "444647_ndsm_1m.tif"
},
{
"sha256": "50730840bf040e4bb24298ade163471f3d0e49a21f0019da79edf07e05618854",
"file": "444648_ndsm_1m.tif"
},
{
"sha256": "dadaeb78e7498f5a22e3750a4aca14d0753cb76cca842294d0b13b775d2dbcd8",
"file": "444649_ndsm_1m.tif"
},
{
"sha256": "af4c91a4fcf86a0070b5a63edfdb98c62ce9a5681f1680cb2c25e08a97330fe6",
"file": "444650_ndsm_1m.tif"
},
{
"sha256": "0cb87bfa286678fb860ce1291be23a5e00f643b2b1f4f0fb1212af44c4ce17d5",
"file": "444651_ndsm_1m.tif"
},
{
"sha256": "f4ef3fb1e2b8e5640f3e597897cbf6b45d08581e79395413de26c43c3df23028",
"file": "445648_ndsm_1m.tif"
},
{
"sha256": "72c3076de6e142f7874f294b65a2c7ed8e9006c7547e3742f70ccb7b0bdea1ec",
"file": "445649_ndsm_1m.tif"
},
{
"sha256": "88d8d647d305f0e82655fab7f745b95b3559cd4cf9ec78292047aca26032d805",
"file": "445650_ndsm_1m.tif"
},
{
"sha256": "7d96ad9ba3acff31cd33d32ce1f95e54d9e2c9668901814b9794871556b9bbd8",
"file": "445651_ndsm_1m.tif"
},
{
"sha256": "309134e88cbe36ded215b8e8e0e39bb0eb6af18ebb5fb0684ee626b515052dcb",
"file": "54442_dtm_1m.tif"
},
{
"sha256": "86cf02e6c0996ddb412cf6580ec8f362a7f8951363a3535905119fa1a0d264ad",
"file": "54451_dtm_1m.tif"
},
{
"sha256": "24c7982c049c596a9d459bdfa15f450fd28e1b75916a47fa397f33322636cfb6",
"file": "54444_dtm_1m.tif"
},
{
"sha256": "b01c3a16e0185c5f444034d43e0b92a5badfc495274378369b8df4b4101d4de1",
"file": "54453_dtm_1m.tif"
}
]
},
"elva": {
"grid_shape": [
2759,
2741
],
"pixel_m": [
1.9996351696461145,
1.999869517941283
],
"canopy_corridor_coverage": 1.0,
"observed_height_percentiles_m": [
5.0703125,
14.0546875,
23.109375,
28.328125
],
"unknown_height_cells": 2016919,
"source_feature_candidates": {
"building": 296,
"water": 63
},
"ndsm_tiles": 29,
"crs": 3301,
"sources": [
{},
{},
{},
{},
{},
{},
{},
{
"sha256": "13948feda38daa1b1804626298c993ff43d0b2e489c4782fae2e3ddc126bce45",
"file": "458641_ndsm_1m.tif"
},
{
"sha256": "c05d6ff79da6b8fb4f9807cd70674994931a4934de2e7f0861c9bdc331c63cc8",
"file": "458642_ndsm_1m.tif"
},
{
"sha256": "e3fb320bf6709b4f69da3eff7f9bb461c313cf6b8ec6f7bfa58a808bd5ce7431",
"file": "459640_ndsm_1m.tif"
},
{
"sha256": "a0e1dceb6c37e504ab59dfeea5980a06b82e64a86e9865462d24f2aafdf753cb",
"file": "459641_ndsm_1m.tif"
},
{
"sha256": "df8855e69c1270af32cf34b8a5ec41460e48404643eea6506a203895437a60d3",
"file": "459642_ndsm_1m.tif"
},
{
"sha256": "fb813815b956602fe491a8482da722c68568f44e03ccc2e4bdf61965e77490cd",
"file": "460640_ndsm_1m.tif"
},
{
"sha256": "61cf44811ef7beafb2d3683e390f92c2b8ec57ac75ff00531835fc8488b459f3",
"file": "460641_ndsm_1m.tif"
},
{
"sha256": "088a2dbbed363cac0923b03a7057bb71048aa72dff747b6f2cf363fc28444b07",
"file": "460642_ndsm_1m.tif"
},
{
"sha256": "534170510b0e723241333f6a949ae7daeef247bac563d1ecb408a8fa71560905",
"file": "460643_ndsm_1m.tif"
},
{
"sha256": "1970ed73244d32a14480eb255e1bb21618978251487dd4802e826b5212629b3f",
"file": "461637_ndsm_1m.tif"
},
{
"sha256": "31e86565c93883fe40597366be607f2697b6c6bbf2805e2a2868dbab26b7897c",
"file": "461638_ndsm_1m.tif"
},
{
"sha256": "83d041606c65c4e4957a509a2c85a71421d08c3dbe9c5692706603bc9d0fdb35",
"file": "461639_ndsm_1m.tif"
},
{
"sha256": "d8cd493ddf31d2943c81d7584e4f5fd2f63fb61754eca8312d290512c1b9646a",
"file": "461641_ndsm_1m.tif"
},
{
"sha256": "10ade7b60b73f9ae488fef6dfef0d9b94541a668e387a1d7a5ca54aea23b9d56",
"file": "461642_ndsm_1m.tif"
},
{
"sha256": "5b27cb1d68d9eb4312372d33b52c05087dafd00a68cfdafc7df93ddf32e088ae",
"file": "461643_ndsm_1m.tif"
},
{
"sha256": "cac25de909c4beca3913b1ea69b63b1bbe314ba3d45a1dc4e23616e823104c06",
"file": "462637_ndsm_1m.tif"
},
{
"sha256": "76ddba0c46943509a933347b5c086f846fb34d33bf1f5e7bfaffe7a3435c786f",
"file": "462638_ndsm_1m.tif"
},
{
"sha256": "09270c0607a35ba1f33f8e54e5a14f51de02149b9e21a13f13d24c205c826c44",
"file": "462639_ndsm_1m.tif"
},
{
"sha256": "64f5c24f38fa748b4ea8b9a01301692d872538156e523a5e0d8578c0bf2cd9ad",
"file": "462640_ndsm_1m.tif"
},
{
"sha256": "1a25fd85920ef9852ee0da1bff113eb72a0861806103530810dd8c8f39f987ca",
"file": "462641_ndsm_1m.tif"
},
{
"sha256": "e2b88ad5583a914821c583c0ac5bf785a46b0daa0be42e86d2404cabc47b4a12",
"file": "462642_ndsm_1m.tif"
},
{
"sha256": "24c9c8287e1182b119100b156cb4c9a6bf1119825e8410ddc7db4d730bcaa352",
"file": "462643_ndsm_1m.tif"
},
{
"sha256": "cbb37981ad6f9d3ad6470aac5ce36e77a5355359b3188b44994559adb9327e9f",
"file": "463638_ndsm_1m.tif"
},
{
"sha256": "8460449a54df26eba1695a19c83cc4ced82d8838fc9a5a8daaa33952d296f439",
"file": "463639_ndsm_1m.tif"
},
{
"sha256": "af17df0a32cee2a076344ad4d1204ff01829ac7d0a0b611595052ba80685a761",
"file": "463640_ndsm_1m.tif"
},
{
"sha256": "f8a576cc057d3d152e45048d6627f1cba9981ca5384644a6230b2131b97a60e0",
"file": "463641_ndsm_1m.tif"
},
{
"sha256": "f385e1fb58eeacf47dd873ae0b80dfd6311d09fe0ae651e515fa6935dd0017ec",
"file": "463642_ndsm_1m.tif"
},
{
"sha256": "abbd885b66bd7f7f50d6388ff3e132f1b95027d7150fcb8b02abc201a7ca16da",
"file": "464640_ndsm_1m.tif"
},
{
"sha256": "a7dcfd9e0c63920d78b3d40229af2daff04b799f0defc7a74a357062d7379431",
"file": "464641_ndsm_1m.tif"
},
{
"sha256": "51a59bc01a5411b46164b38161ccfc32b7670177d9f520e25610eb34e29d40f5",
"file": "54534_dtm_1m.tif"
},
{
"sha256": "ac565598c564068b4570db031416caa168309f50ea7b9af96b123727e64cb3c3",
"file": "54543_dtm_1m.tif"
},
{
"sha256": "a2c5085103ed67b76bfa3dfbf1645144091efe358fc2cf5dbeb47e5cdcd39f8b",
"file": "54632_dtm_1m.tif"
},
{
"sha256": "8c179cbb19c2d7ff6cd6e67aff7f56d22ffe644048732a2236da7b536710e74f",
"file": "54641_dtm_1m.tif"
}
]
},
"otepaa": {
"grid_shape": [
4476,
4540
],
"pixel_m": [
1.9997004405286343,
1.9997721179624663
],
"canopy_corridor_coverage": 1.0,
"observed_height_percentiles_m": [
4.9765625,
14.15625,
23.296875,
27.921875
],
"unknown_height_cells": 11853928,
"source_feature_candidates": {
"building": 291,
"water": 103
},
"ndsm_tiles": 36,
"crs": 3301,
"sources": [
{},
{},
{},
{},
{},
{},
{},
{
"sha256": "2e5407be951f47c3567ce265451ec9ab5061c143b860a4a37279c1e2f4ce95be",
"file": "426653_ndsm_1m.tif"
},
{
"sha256": "5af45abbfcec2109599a78b85d9f35a9fc9558cb0aaa9aa12cd8074749b4fbe9",
"file": "426654_ndsm_1m.tif"
},
{
"sha256": "becb6a5bbb43362688732b88a523d746026a309387da04f1edd5a4672cc2bf8a",
"file": "426655_ndsm_1m.tif"
},
{
"sha256": "06f1877c9435574c674a7099598f56273fee90497f051735aaef56fcdb361ee2",
"file": "426656_ndsm_1m.tif"
},
{
"sha256": "bbbfaa24b5f2de2abdabaf783d39c2848c95175739cea575dbe45b136f76da01",
"file": "427651_ndsm_1m.tif"
},
{
"sha256": "49ee19dba71da97b31d5caa2a94e911d9a2d7e65affacc229ca84338a68582c8",
"file": "427652_ndsm_1m.tif"
},
{
"sha256": "9ace8b790469898793832dc8d17b9a8ac66b18d87339c3f8c68cd3d693ef5497",
"file": "427653_ndsm_1m.tif"
},
{
"sha256": "d4999e926d875b83a954262563bd83f765f0b181161b56513d73fe720b816680",
"file": "427654_ndsm_1m.tif"
},
{
"sha256": "c1fb132b2340ad890c13964d92f9c5f6bb5e2fae03190479afe7225adf16f8af",
"file": "427655_ndsm_1m.tif"
},
{
"sha256": "4792486bbef03ba6d0c7e6d5e457a613428257414668b565fb76573ab04acfee",
"file": "427656_ndsm_1m.tif"
},
{
"sha256": "1dd326a610fc57faa4af0296074807e355e778b4cc64b97a0946239ab5c6d53f",
"file": "428650_ndsm_1m.tif"
},
{
"sha256": "23a37a5274114fec8088ef7cee9996b4a84083e933d3b1e5dc795ce7043e0cb7",
"file": "428651_ndsm_1m.tif"
},
{
"sha256": "5c95c2c1919a536df585b696fe9cb34aa08710b5677486a074b870771819676c",
"file": "428652_ndsm_1m.tif"
},
{
"sha256": "91ee226d378db91ce685734276e398ec59cc38b52654924ed0cced52ed156626",
"file": "428653_ndsm_1m.tif"
},
{
"sha256": "bb6e4291ede6b4faf401bbe1b8d8b79c7d23986863a792dff29fa268f604a2bf",
"file": "428655_ndsm_1m.tif"
},
{
"sha256": "c15ed8f87c340d561ebfacab786682eb94bca75e815db556dbb89f7d371f81b9",
"file": "428656_ndsm_1m.tif"
},
{
"sha256": "d98539a488ebba81770042dc5736aebd971f2e90ede99830f192f935f4ea8b09",
"file": "429650_ndsm_1m.tif"
},
{
"sha256": "43280022aca994561eebb696ebf45ec161250c16226e309a90313fa616290ec5",
"file": "429651_ndsm_1m.tif"
},
{
"sha256": "0bd169d6c0c7308657d10418b611508945fee39558689d2f69973aee4b14de57",
"file": "429652_ndsm_1m.tif"
},
{
"sha256": "611ee6eb97628d2f18ec7771113b7c4382d0d8a87bc9e627e103e54e695f9ea0",
"file": "430650_ndsm_1m.tif"
},
{
"sha256": "20f7f349cf5fa47c0bf313f2f62236d1430e1947047486439892ce5509f309ad",
"file": "430651_ndsm_1m.tif"
},
{
"sha256": "ab5940319afb62ee7a25fcb73df8ced7b5298226747c0bb0e17e0e6735e122e0",
"file": "430652_ndsm_1m.tif"
},
{
"sha256": "b3ba1a7f173d11d8615441d6d0c06c4436a0fcbeeabe78491bfcd8c1938e81d9",
"file": "431650_ndsm_1m.tif"
},
{
"sha256": "169b8fa5df1ebbb6540fe1dc0b605f16a3958ec12ab478e30f98c6b52dbc3898",
"file": "431651_ndsm_1m.tif"
},
{
"sha256": "94084377c0df7f90ed221974bc186607ac27c4bda05f9a78a5b8fb9ccd7d37b9",
"file": "431652_ndsm_1m.tif"
},
{
"sha256": "053874f15a67c55c8515ac30505b5bb0f9baf181ce31157b5f086f000d0daa30",
"file": "432650_ndsm_1m.tif"
},
{
"sha256": "b28de529c6833104b445e9cebdd5d6ca9b1e13a3db9a438523757fa96de113ed",
"file": "432651_ndsm_1m.tif"
},
{
"sha256": "19fef4b89e6835df362b9d6634b2f4b300f79c6e6e043dd5a40fefcfc9591ac2",
"file": "433647_ndsm_1m.tif"
},
{
"sha256": "342daf0a97fcc0b1e506009aec3b206abe0506ea811b8134332a63b1d024095a",
"file": "433648_ndsm_1m.tif"
},
{
"sha256": "9fde79821ca468040cd0305d063fc70004c7698ffdd3d3b979766fde3f7edccb",
"file": "433649_ndsm_1m.tif"
},
{
"sha256": "00922706c3f22ae9979ddc700b685a6de42c920ace4af726c7e01c2f9c8ebac9",
"file": "433650_ndsm_1m.tif"
},
{
"sha256": "049c86456fe455b94b8f824f5052b8033532391c57c74a595711ece129f8b150",
"file": "433651_ndsm_1m.tif"
},
{
"sha256": "badbdf598b8921e3c9529820f40271c8a4009e0b8816c91e3d469d5838fa1cef",
"file": "434647_ndsm_1m.tif"
},
{
"sha256": "98999bf542effbe530ccea2c81c42407fa6baa9832cdd4780ee20306a1084b1c",
"file": "434648_ndsm_1m.tif"
},
{
"sha256": "32e069fd4f29a9fd40ed7d10ce5191677265dbe80a2b53bcaca4d5ce817b38a4",
"file": "434649_ndsm_1m.tif"
},
{
"sha256": "8819419a724ec0dfbf05ece4722e43b1a305794bec6eef632a71c17af19cfdfe",
"file": "434650_ndsm_1m.tif"
},
{
"sha256": "19e133fa04a40783a27ea20d102694767d1f128e0a353433a8e8f123bc1233fd",
"file": "54244_dtm_1m.tif"
},
{
"sha256": "74ffa407e070c143167421ee923c6f3fa01cc42056caaf1081a615cafaffbba9",
"file": "54253_dtm_1m.tif"
},
{
"sha256": "c84ee474243f8d6cd51bd1cca9fbfaf8e02fa070c1e50ea3fb570f97c3dae84c",
"file": "54254_dtm_1m.tif"
},
{
"sha256": "95c90bba89ccee813ec1e35e4809c96cc927af6ece6a6bfa15a63f9bad0ec958",
"file": "54342_dtm_1m.tif"
},
{
"sha256": "7ac10790028f1e74e78bd47f706993c62426477e2a264047d1d804953bbacc0c",
"file": "54351_dtm_1m.tif"
},
{
"sha256": "af9c2e888c2970e37d5715ab8e9f78dcab3c9b71e2984db25ada46169d799d10",
"file": "54352_dtm_1m.tif"
}
]
}
}{
"resolution": [
1920,
1080
],
"fps": 30,
"frames_per_clip": 150,
"lighting": "golden",
"msaa_samples": 8,
"camera_audits": {
"ruuhimaki": [
{
"mode": "chase",
"samples": 39998,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 0.5850000000000001,
"max_horizontal_acceleration": 14.478967987234059,
"max_horizontal_speed": 51.72873080549647,
"max_vertical_speed": 6.455832030856514,
"max_vertical_acceleration": 3.1841035151132266,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
},
{
"mode": "heli",
"samples": 39998,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 0.9886500000000003,
"max_horizontal_acceleration": 15.333182533931527,
"max_horizontal_speed": 54.828947921383,
"max_vertical_speed": 6.76092226451708,
"max_vertical_acceleration": 3.1578229138103118,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
}
],
"neeruti": [
{
"mode": "chase",
"samples": 37108,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 0.45,
"max_horizontal_acceleration": 17.02817815281118,
"max_horizontal_speed": 50.380777859092895,
"max_vertical_speed": 4.522570644671464,
"max_vertical_acceleration": 3.1002396056042736,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
},
{
"mode": "heli",
"samples": 37108,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 1.2852450000000004,
"max_horizontal_acceleration": 13.013383071547697,
"max_horizontal_speed": 52.00983300790811,
"max_vertical_speed": 6.592610203746105,
"max_vertical_acceleration": 3.5629840169869804,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
}
],
"elva": [
{
"mode": "chase",
"samples": 45757,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 0.5850000000000001,
"max_horizontal_acceleration": 16.710377522178927,
"max_horizontal_speed": 52.29048664028828,
"max_vertical_speed": 5.6055807241970195,
"max_vertical_acceleration": 3.286984702804116,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
},
{
"mode": "heli",
"samples": 45757,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 1.2852450000000004,
"max_horizontal_acceleration": 14.137438785105704,
"max_horizontal_speed": 53.885486428673104,
"max_vertical_speed": 6.0311892772377895,
"max_vertical_acceleration": 3.163673460977745,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
}
],
"otepaa": [
{
"mode": "chase",
"samples": 64039,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 0.5850000000000001,
"max_horizontal_acceleration": 15.613726108860865,
"max_horizontal_speed": 54.60469046131732,
"max_vertical_speed": 6.0433121308510165,
"max_vertical_acceleration": 3.8259314198200522,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
},
{
"mode": "heli",
"samples": 64039,
"proxy_intrusions": 0,
"inter_sample_clearance_lift": 0.0,
"horizontal_smoothing_seconds": 0.9886500000000003,
"max_horizontal_acceleration": 15.985896676613176,
"max_horizontal_speed": 56.240572600139686,
"max_vertical_speed": 6.011052591800435,
"max_vertical_acceleration": 3.5867489398957586,
"clearance_model": "Wind-inflated crowns, static scenery/building bounds and terrain; not a triangle-level collision proof."
}
]
},
"timing_scope": "Wall time around direct HDR render and blocking BGR readback; not streamed FPS.",
"temporal_metric": "Quarter-resolution bidirectional-flow-consistent luminance residual; diagnostic only.",
"stages": {
"ruuhimaki": {
"trees": 517911,
"species": {
"0": 125890,
"1": 90849,
"2": 90537,
"3": 143565,
"4": 459,
"5": 490,
"6": 36929,
"7": 29192
},
"seconds": 333.31,
"mapped_buildings": 232,
"water_bodies": 1
},
"neeruti": {
"trees": 315013,
"species": {
"0": 103562,
"1": 47025,
"2": 46845,
"3": 52826,
"4": 3345,
"5": 3735,
"6": 37874,
"7": 19801
},
"seconds": 309.23,
"mapped_buildings": 94,
"water_bodies": 33
},
"elva": {
"trees": 423672,
"species": {
"0": 130297,
"1": 59302,
"2": 59529,
"3": 70870,
"4": 4226,
"5": 4522,
"6": 61699,
"7": 33227
},
"seconds": 381.3,
"mapped_buildings": 270,
"water_bodies": 63
},
"otepaa": {
"trees": 651447,
"species": {
"0": 203067,
"1": 92825,
"2": 92888,
"3": 108290,
"4": 7402,
"5": 8264,
"6": 90381,
"7": 48330
},
"seconds": 533.65,
"mapped_buildings": 251,
"water_bodies": 103
}
},
"public_originals": "All original MP4 clips and linked PNGs are hosted on this site."
}Report previews preserve composition and colour but are resized/compressed for reading. They are not substitutes for the linked lossless stills when inspecting subpixel foliage. Video playback is user-initiated at the original rate. The evidence supports the stated finite acceptance corpus, not a claim that every possible scene, view or future dataset is artifact-free.