EdgeRecon — reconstruction progress

Real driving footage next to the reconstructed scene, with the numbers. Every review is a recorded pipeline run.

This page records EdgeRecon's reconstruction pipeline on real driving data: multi-camera + LiDAR footage in, a Gaussian-splat scene out (static background, tracked vehicles placed back as animated actors, OpenUSD for Isaac Sim), then variations of that scene: weather and light, and a pothole carved into the road and driven over with vehicle physics.

LIVE: everything shown here ran end to end in the pipeline. Each review shows the real camera frames beside the render, how every score is computed, and what still looks wrong. ROADMAP: EdgeRecon's own Indian-road captures. The clips here come from NVIDIA's PhysicalAI Autonomous Vehicles dataset (US roads, 7 cameras + roof LiDAR), not from our own rig.

The pipeline is built on NVIDIA's open-sourced tooling (NCore, Instant NuRec and NuRec, Asset Harvester, 3DGRUT). The vehicles are not reconstructed: a diffusion model generates them from a few camera crops and they are placed on their LiDAR tracks, so a car can look plausible and still differ from the real one.

Each clip has one review page. It shows the fast Instant NuRec background and the full NuRec training background side by side once both have rendered the same frames, plus every scene variation made from that clip. Large 3D files (USD scenes, splat PLYs) are not hosted here; they are shared on request.

Reviews

US intersection, about 20 s 9fc15f8c

One clip from NVIDIA's PhysicalAI Autonomous Vehicles dataset: 7 cameras and a roof LiDAR, driving through an intersection.

Open the review →

Instant NuRec: front camera masked PSNR 19.42 dB at coverage 0.42, 15 vehicles placed back · run 2026-10-29 14:40 UTC
Full NuRec: not on this page yet.
Variations: heavy rain, dust haze, dusk, night with headlights, pothole with vehicle physics, pothole + heavy rain

Earlier runs of this clip: 2026-10-29 09:41 UTC (Instant NuRec).

Milestones

2026-09-18 — First end-to-end run of the unified pipeline. One command from the raw clip to a validated USD scene, every stage on real data.
2026-09-24 — Rebuilt from code on one Python environment. Every stage reinstalled from the repository into a single environment and run end to end; one GPU-specific fix (an attention-kernel dispatch) was needed.
2026-10-11 — Full NuRec training mode end to end, review page. 40,000-step NuRec background composed with the actors; the first side-by-side review of real cameras and renders. A 23 m frame offset in full mode was found and fixed on the way.
2026-10-28 — Actors face the right way and appear only while observed. 12 of 13 measurable actors had been placed backwards; fixed and checked against the real cameras. Scores now count only the pixels the render draws, with coverage shown next to them.
2026-10-29 — Scene variations with physics. Rain, dust, dusk and night applied to the reconstructed clip (labels unchanged); a ground mesh from the accumulated LiDAR (2.7 cm median agreement with the reconstructed road); a pothole carved into mesh and scene, driven over by a simulated vehicle whose body motion moves the camera, also in Isaac Sim. One clip so far. See the run.
2026-10-29 — Rebuilt from code alone, same result. The machine was taken back to code only (environment, data and download caches removed), rebuilt from the repository, and the clip run again: the same background, ground mesh, pothole and physics numbers; front-camera masked PSNR 19.42 dB against 19.40 dB before. See the review.