| Mode | PSNR s | SSIM s | LPIPS s | PSNR full | SSIM full | LPIPS full | Actors | Static Gaussians |
|---|---|---|---|---|---|---|---|---|
| Instant NuRec | 12.52 | 0.367 | 0.572 | 12.53 | 0.365 | 0.592 | 17 | 1,912,472 |
| Full NuRec | 13.67 | 0.523 | 0.573 | 13.73 | 0.520 | 0.590 | 17 | 2,710,811 |
Front camera every frame; the other cameras every 2 frames (render budget: 0.33 s/frame measured on the first 20 front frames, 22.0 min per mode). Boxes: cyan = composed actor (labelled with its track id), thin grey = detected by the LiDAR detector but not composed (no usable Step 1 crop, Step 2 failed, or rejected by the QA gate). Space plays/pauses, ←/→ steps one frame, F flips A/B.
| Camera | Frames | Instant NuRec | Full NuRec | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PSNR s | SSIM s | LPIPS s | PSNR | SSIM | LPIPS | PSNR s | SSIM s | LPIPS s | PSNR | SSIM | LPIPS | ||
| Front wide 120° | 599 | 12.52 | 0.367 | 0.572 | 12.53 | 0.365 | 0.592 | 13.67 | 0.523 | 0.573 | 13.73 | 0.520 | 0.590 |
| Cross left 120° | 300 | 10.62 | 0.296 | 0.671 | 10.75 | 0.298 | 0.703 | 13.47 | 0.534 | 0.528 | 13.54 | 0.528 | 0.558 |
| Cross right 120° | 300 | 12.88 | 0.324 | 0.652 | 12.87 | 0.321 | 0.681 | 15.47 | 0.567 | 0.521 | 15.44 | 0.562 | 0.544 |
| Rear left 70° | 300 | 9.13 | 0.227 | 0.792 | 9.26 | 0.228 | 0.830 | 14.81 | 0.661 | 0.492 | 14.84 | 0.651 | 0.521 |
| Rear right 70° | 300 | 9.76 | 0.221 | 0.803 | 9.84 | 0.218 | 0.835 | 14.06 | 0.614 | 0.508 | 14.06 | 0.606 | 0.533 |
| Front tele 30° | 300 | 17.00 | 0.478 | 0.623 | 15.97 | 0.460 | 0.684 | 16.99 | 0.566 | 0.507 | 16.75 | 0.556 | 0.550 |
"s" = static-only (ego vehicle + every detected box masked). PSNR over unmasked pixels; SSIM = mean of the SSIM map over unmasked pixels; LPIPS (AlexNet) after filling masked pixels of the render with the real ones. 6 cameras, 960 px wide.
Each render uses the camera's own calibration and its rolling-shutter start/end pose at that frame. Orange = edges in the real frame, cyan = edges in the render, white = both. The shift is the sub-pixel phase-correlation offset between the two frames (ego vehicle masked).
The same drive with the camera moved 1.5 m and 3.0 m to the left (rig +y). There is no real footage for these views, since the car never drove there. This is the novel-view test: the scene has to hold up away from the path it was trained on. Instant NuRec: a point 15 m ahead moved from u=483.3 px to u=535.6 px for 1.5 m left — the expected direction. Full NuRec: a point 15 m ahead moved from u=483.3 px to u=535.6 px for 1.5 m left — the expected direction.
For each mode the scene was composed again without the actor that covers the most of the front camera (into a scratch directory; the real scene is untouched), and rendered at the frames where that car was closest. If the background behind it was reconstructed, the right image shows road; if not, a hole or a smear.
Left: real Step 1 crops the diffusion model saw. Right: the generated asset, fitted to the LiDAR cuboid, from the front, three-quarter, side and rear.
Raw scene files (usdz/PLY, 2.15 GB) available on request.
| File (not hosted) | MB | What |
|---|---|---|
| review_assets/instant/scene_instant.usdz | 632.9 | Composed scene (instant background + actors), OpenUSD, for Isaac Sim |
| review_assets/instant/static_scene_instant.ply | 147.3 | Static background only (instant), 3DGS PLY, SH degree 0 |
| review_assets/instant/combined_busiest_instant.ply | 176.7 | Static + actors at the busiest instant (instant), 3DGS PLY, for SuperSplat |
| review_assets/full/scene_full.usdz | 781.0 | Composed scene (full background + actors), OpenUSD, for Isaac Sim |
| review_assets/full/static_scene_full.ply | 184.3 | Static background only (full), 3DGS PLY, SH degree 0 |
| review_assets/full/combined_busiest_full.ply | 231.0 | Static + actors at the busiest instant (full), 3DGS PLY, for SuperSplat |
SuperSplat (superspl.at/editor) opens the PLYs in a browser. It could not be tested headless here; instead the combined PLY was read back with plyfile. Instant NuRec: 2,599,045 Gaussians at t=0.0 s with 7 actors; standard 3DGS properties present: yes (checked with plyfile). Full NuRec: 3,397,384 Gaussians at t=0.0 s with 7 actors; standard 3DGS properties present: yes (checked with plyfile).
Input: one clip from NVIDIA's PhysicalAI Autonomous Vehicles dataset, with 7 cameras and a roof LiDAR, about 20 s at an intersection. These are not Indian roads and not EdgeRecon's own rig.
live Raw clip → NCore conversion · aux masking/depth · Instant NuRec · full NuRec training (A100) · Asset Harvester · USD composition + validation
roadmap EdgeRecon's own rig and Indian-road capture · scenario variation (Replicator, Cosmos) · NuRec's native dynamic-actor layer (not functional in the public release)
gap close-range LiDAR cuboid fitting in dense traffic (diagnosed, not solved)
Status pills follow EdgeRecon_Master_Status.md §7 and §3; nothing here goes beyond what it records as run.
| Step | Wall (m:s) | Card peak MiB | Runs |
|---|---|---|---|
| 01_detect_all_vehicles_nuscenes | 0:35 | 1,160 | 1 |
| 02_ncore_convert | 2:16 | 507 | 1 |
| 03_aux_container | — | — | 1 |
| 03_reconstruction_full_export | — | — | 3 |
| 03_reconstruction_full_train | — | — | 1 |
| 03_reconstruction_instant | 1:10 | 11,035 | 1 |
| 04_harvest_step1_ncore_parser | 1:45 | 5,674 | 1 |
| 04_harvest_step2 | 17:49 | 20,324 | 1 |
| 07_review_render_full | — | 3,276 | 2 |
| 07_review_render_instant | 5:11 | 2,910 | 1 |
"—": the docker steps' watchdog is stopped before it writes its peak line, and a re-run appends to the same log, so no single wall time or peak can be read from it.