Add nuScenes (v1.0) dataset support

[stepankonev]

Summary

Adds support for the nuScenes dataset (Motional, CVPR 2020) — the de-facto surround-view E2E / BEV benchmark: 1000 ~20 s scenes, a 6-camera surround rig (1600×900), a 32-beam LIDAR_TOP, and densely annotated 3D boxes at 2 Hz keyframes. One keyframe sample → one StandardFrameData; one scene → one segment.

Modalities emitted

Modality	Source
Cameras	6 CAM_* surround views mapped onto the canonical CameraDirection members; downscaled via cameras_identity_adapter max_size.
LiDAR	LIDAR_TOP cloud (xyz, ego frame).
3D detections (vector)	sample_annotation boxes transformed global→ego; category_name folded into the coarse DetectionType.
3D detections (BEV raster)	the same boxes rasterized by Detections3DBEVAdapter on the same grid as the HD-map raster, so the two co-register pixel-for-pixel.
HD map (BEV)	vector map-expansion (arcline lane centers, lane/road dividers, crossings, walkways, stop lines, drivable area, intersections) → unified MapElementType in the ego frame → rasterized by HDMapBEVAdapter.
Past/future trajectory	ego poses, via the segment-context aggregator.

Design notes

• Read directly from the JSON tables — nuscenes-devkit is not a runtime dependency (it pins numpy<2, which conflicts with the project's numpy 2.x). The split scene-lists and the lane-arcline discretization are vendored from the devkit (Apache-2.0).
• --split is an official nuScenes label that also selects the metadata version: mini_train/mini_val → v1.0-mini, train/val → v1.0-trainval, test → v1.0-test. The test split ships no annotations.
• Conversion-rate optimization — only LIDAR_TOP keyframe sample_data rows are loaded (~410 k vs 2.6 M), giving ~32 frames/s at 8 workers.
• Partial trainval converts cleanly — scenes whose sensor blob is not yet on disk are detected and skipped, so conversion works while a download is still in progress.
• HD map is optional — rasterized only when the separate nuScenes-map-expansion-v1.3 pack is unzipped into <dataroot>/maps/; otherwise the map is skipped.

Limitations

• The 5 radars have no StandardE2E target yet and are not ingested.
• Ships as .tgz archives and must be extracted first (scripts/extract_nuscenes.sh, or scripts/prepare_dataset_nuscenes.sh to extract + preprocess in one step).

Tests

• tests/dataset_processors/test_nuscenes_dataset_processor.py — box geometry, global→ego transform, category_name → DetectionType folding, camera-channel mapping, partial-download detection, etc. Real-frame checks are gated on NUSCENES_DATAROOT (default v1.0-mini), so CI without the data still runs the unit-level assertions.
• Box orientation verified against LiDAR PCA (not point-containment, which is blind to 90° yaw errors).

Validation

mini_train converted end-to-end; on real frames I verified box dimensions (ped 0.85 m, vehicle 4.71 m, bike 2.06 m), map alignment (ego ∈ drivable area), and left/right BEV orientation — all correct. Per-scene videos (6-camera grid with projected boxes + BEV from the npz rasters) were rendered from the processed .npz to sanity-check every modality.

[codecov]

Codecov Report

❌ Patch coverage is 88.97959% with 54 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...rc_datasets/nuscenes/nuscenes_dataset_converter.py	36.66%	19 Missing ⚠️
...rc_datasets/nuscenes/nuscenes_dataset_processor.py	87.50%	13 Missing ⚠️
...e2e/caching/src_datasets/nuscenes/_nuscenes_map.py	91.96%	9 Missing ⚠️
..._e2e/caching/src_datasets/nuscenes/_nuscenes_io.py	94.52%	8 Missing ⚠️
...caching/src_datasets/nuscenes/_nuscenes_arcline.py	95.23%	4 Missing ⚠️
standard_e2e/caching/process_source_dataset.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!