METEOR demo scenes for Autoware (meteor-demo-scenes)

Six short driving scenes, one per road type (147–148 frames each, 8 synchronised cameras, ego motion, LiDAR raster) to run the released METEOR model (AutowareFoundation/meteor) and the demo renderers of https://github.com/tier4/METEOR without access to the training corpus. These are the exact scene roots the Orin demos and benchmarks in the repository refer to (valday, valcurve, fast).

The scenes come from the validation split (a held-out recording day) of the METEOR corpus — TIER IV Co-MLOps Data Recording System (DRS) recordings on public roads in Japan, 2026-01-23. They are not training scenes. No ground truth is included; the model runs GT-free.

Root	Scenes	Frames	Content	Size
highway_day/	1	147	daytime elevated expressway, straight, ~90 km/h	110 MB
mountain_day/	1	147	daytime winding mountain road, guard rails, forest	137 MB
arterial_day/	1	148	daytime multi-lane urban arterial, straight-ahead	132 MB
valday/	1	147	daytime urban, intersection and curves (the default demo root)	121 MB
valcurve/	1	148	night, elevated expressway with tight curves	84 MB
fast/	1	147	dusk expressway, used for --bench latency runs	88 MB

arterial_day, valday, valcurve and fast come from the validation split (a held-out recording day, 2026-01-23); highway_day and mountain_day come from training recordings (October 2025) because the held-out day has no daytime expressway or mountain driving. The model has seen those two roots during training; treat them as illustration, not as a test set.

Anonymisation

Every image was regenerated from the anonymised camera stream of the source dataset, in which faces and licence plates are blurred before any use. manifest.json carries "anonymized": 1. No un-anonymised pixel is part of this repository.

Layout

<root>/
├── scenes.txt                          scene names, one per line
└── <scene>/
    ├── manifest.json                   camera intrinsics / extrinsics + per-frame file table
    ├── img/<frame>_<CAM_NAME>.jpg      768×432 JPEG, 8 cameras per frame (anonymised)
    ├── ego_motion.npz                  v0 (speed, m/s), pose (x, y, yaw), future waypoints, controls
    └── lidar_bev/<frame>.npz           key "lb": float16 pillar raster [4, 400, 250] @ 0.4 m (optional model input, cast to float32)

Camera names: CAM_FRONT_WIDE, CAM_FRONT_LEFT, CAM_FRONT_RIGHT, CAM_BACK_WIDE, CAM_BACK_LEFT, CAM_BACK_RIGHT, CAM_FRONT_NARROW, CAM_BACK_NARROW.

manifest.json keys that matter for inference: cams[<CAM>].K (3×3 at 768×432), cams[<CAM>].T_ego_cam (4×4, camera → ego; the model takes its inverse), frames[i].imgs[<CAM>], frames[i].lidar_bev. The manifests also list the ground-truth files of the full training format (gt/, occ/, seg2d/, …); those directories are intentionally not shipped here and the runtimes do not read them.

Use

pip install -U huggingface_hub
hf download AutowareFoundation/meteor-demo-scenes --repo-type dataset --local-dir demo

# onnxruntime, one frame (repo: https://github.com/tier4/METEOR)
python3 hf/onnx_smoke_test.py --onnx meteor_v157c3Z.onnx --root demo/valday --frame 40

# Jetson AGX Orin, real-time demo (C++ runtime + the released INT8 engine)
meteor_realtime --engine eng/v157c3Zg_int8.engine --root demo/valday --out out/demo.mp4
meteor_realtime --engine eng/v157c3Zg_int8.engine --root demo/fast --bench 40

The Python and C++ runtimes in the repository (deploy/orin_realtime.py, deploy/cpp) take any of these roots via --root.

Notes

• Public-road imagery: faces and licence plates are not blurred. Use for evaluating and demonstrating the model only; do not redistribute frames out of this context.
• Ego speed comes from the vehicle's own odometry; the LiDAR raster is the same pillar raster used for the optional LiDAR input of the model (zeros when the input is not available).
• SHA256SUMS covers every file.

License

Follows the license of the METEOR repository, which has not been finalised yet; until then this data is provided for research and demonstration use only.