Apppublic

AutowareFoundation/meteor-multitask-driving-demo

sourceHugging Faceapache-2.0updated 9d agoView on Hugging Face

1likes

App README

METEOR — surround-view multi-task driving perception

One forward pass of a 54 M-parameter network over 8 synchronised cameras produces 12 outputs simultaneously:


BEV lane / road map (800×500 @ 0.2 m)	3D boxes + parked-vehicle state
per-camera metric depth (64 bins)	21-class 2D semantic segmentation
multi-scale 2D detections (10 classes)	3D semantic occupancy (10 classes)
per-agent 3-s trajectories	BEV collision-risk map
traffic-light state	ego driving plan (3 modes, 6 waypoints)

The demo runs the released ONNX graph on short clips of the authors' own recordings and renders every head into a single video, reproducing the upstream reference visualiser.

Attribution

—Model: `AutowareFoundation/meteor` (Apache-2.0)
—Scenes: `AutowareFoundation/meteor-demo-scenes`, recorded by TIER IV (Co-MLOps DRS), faces and licence plates blurred. Frames are streamed from the dataset at runtime and are not redistributed here.
—Post-processing (meteor_decode.py) and visualisation (meteor_render.py, meteor_viz_np.py, meteor_occ_iso.py) are ports of deploy/runtime.py, deploy/orin_render.py, deploy/viz_np.py and deploy/occ_iso.py from tier4/METEOR (Apache-2.0). The raw ONNX graph emits only tensors — the box decode, NMS, lane thinning, temporal fusion and occupancy rendering all live outside it.

Notes

—The network is single-frame (its temporal memory was baked out for export). Only the BEV lane map, the box yaw and the ego path ribbon carry state across frames, via the post-processing in this Space.
—Camera order is fixed by the graph: CAM_FRONT_WIDE, CAM_FRONT_LEFT, CAM_FRONT_RIGHT, CAM_BACK_WIDE, CAM_BACK_LEFT, CAM_BACK_RIGHT, CAM_FRONT_NARROW, CAM_BACK_NARROW.
—Research artefact: the released weights must not be used to control a vehicle.