How gyuz1/e2e_ad_cell-planner works
Cell Planner is an end-to-end camera driving planner for the ETRI challenge. Its main workflow loads multi-frame, multi-camera scenes, builds BEV features, estimates ego status from vision, generates trajectories for command-conditioned spatial cells, then uses the target point to select a candidate. Training supports no-distillation and teacher/student distillation variants; evaluation measures trajectory quality and latency, while submission tooling packages inference results. The README documents these capabilities; some links to unsampled training and submission internals are therefore documented rather than source-verified.
Main entry points: VAD_head.py VAD planning head · VAD.py VAD detector · nuscenes_vad_dataset.py VAD dataset · eval_holdout_l2_and_tinfer.py Holdout evaluation
Components 19
Data and geometry
- VAD dataset
projects/mmdet3d_plugin/datasets/nuscenes_vad_dataset.py - Image transforms
projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py - Geometry cache
tools/cache/etri_geometry_cache.py
BEV perception
- VoVNet backbone
projects/mmdet3d_plugin/models/backbones/vovnet.py - VAD detector
projects/mmdet3d_plugin/VAD/VAD.py
Planning and distillation
- VAD planning head
projects/mmdet3d_plugin/VAD/VAD_head.py - Cell generation and routing
projects/mmdet3d_plugin/VAD/cell_planner_utils.py - Goal anchors
projects/mmdet3d_plugin/VAD/goal_anchor_utils.py - Trajectory decoder
projects/mmdet3d_plugin/VAD/modules/decoder.py - LAW world model
projects/mmdet3d_plugin/LAW/bev_latent_world_model.py - VAD-LAW integration
projects/mmdet3d_plugin/LAW/VAD_LAW.py - Trajectory distillation
projects/mmdet3d_plugin/LAW/trajectory_distillation.py
Train and evaluate
- Training configurations
configs/kd_delta.py - Training entry point
tools/train.py - Holdout evaluation
tools/eval_holdout_l2_and_tinfer.py - Challenge submission
tools/etri_test_submit.py
Other components
- Driving scenes
- Selected trajectory
- Privileged teacher
Connections 21
- Driving scenes → VAD dataset: loads scenes
README.md - VAD dataset → Image transforms: prepares images
- Image transforms → Geometry cache: uses cached geometry
- Image transforms → VoVNet backbone: provides images
- VoVNet backbone → VAD detector: extracts features
- VAD detector → VAD planning head: passes BEV features
- VAD planning head → Trajectory decoder: decodes trajectories
- VAD planning head → Cell generation and routing: generates and routes
projects/mmdet3d_plugin/VAD/VAD_head.py - VAD planning head → Goal anchors: uses goal anchors
projects/mmdet3d_plugin/VAD/VAD_head.py - VAD planning head → Selected trajectory: returns selected path
README.md - VAD-LAW integration → LAW world model: computes world loss
projects/mmdet3d_plugin/LAW/VAD_LAW.py - VAD-LAW integration → Trajectory distillation: applies distillation
projects/mmdet3d_plugin/LAW/VAD_LAW.py - Privileged teacher → Trajectory distillation: provides features
README.md - Training configurations → Training entry point: configures training
README.md - Training entry point → VAD dataset: trains on data
- Training entry point → VAD detector: optimizes model
- Holdout evaluation → VAD dataset: replays windows
tools/eval_holdout_l2_and_tinfer.py - Holdout evaluation → VAD detector: runs inference
- Holdout evaluation → Selected trajectory: scores trajectories
README.md - Challenge submission → VAD detector: runs challenge model
- Challenge submission → Selected trajectory: writes submission
README.md