Skip to content

Usage

This page documents the command surface that is visible in the repository today. It does not invent an environment lockfile or runtime benchmark setup that the repository does not contain.

Installation assumptions

The repository does not ship a requirements.txt, environment.yml, or other pinned environment manifest. The checked-in README and configs indicate the following stack assumptions:

  • Python with CUDA-enabled PyTorch
  • mmcv-full 1.7.x
  • mmdet 2.28.x
  • mmdet3d 1.0.0rc6
  • nuscenes-devkit

Set the repository root on PYTHONPATH before invoking the tools:

export PYTHONPATH=.

Dataset layout

The active nuScenes config expects the dataset under data/nuscenes/ and references temporal annotation files:

  • data/nuscenes/nuscenes_infos_temporal_train.pkl
  • data/nuscenes/nuscenes_infos_temporal_val.pkl

tools/create_data.py contains the repository's nuScenes conversion path and writes temporal info files through nuscenes_converter.create_nuscenes_infos(...). The script requires the MMDetection3D stack to be installed before it can run.

Example conversion command shape:

python3 tools/create_data.py nuscenes \
  --root-path data/nuscenes \
  --canbus data \
  --version v1.0 \
  --max-sweeps 10

Train

python3 tools/train.py projects/configs/bevformer/bevformer_project.py

To resume from a checkpoint:

python3 tools/train.py \
  projects/configs/bevformer/bevformer_project.py \
  --resume-from work_dirs/bevformer_project/iter_100000.pth

Evaluate

python3 tools/test.py \
  projects/configs/bevformer/bevformer_project.py \
  work_dirs/bevformer_project/iter_100000.pth \
  --eval bbox

Visualize BEV outputs

python3 tools/test.py \
  projects/configs/bevformer/bevformer_project.py \
  work_dirs/bevformer_project/iter_100000.pth \
  --eval bbox \
  --viz-bev --viz-num 20 --viz-score-thr 0.2 \
  --viz-outdir work_dirs/bevformer_project/bev_viz

Benchmark command surface

The repository includes a dedicated benchmark entry point in tools/analysis_tools/benchmark.py. It requires the full CUDA/MMCV/MMDetection3D stack and a real checkpoint.

python3 tools/analysis_tools/benchmark.py \
  projects/configs/bevformer/bevformer_project.py \
  --checkpoint work_dirs/bevformer_project/iter_100000.pth \
  --samples 200

Result probes

The repository includes two post-processing probes for nuScenes result files:

python3 tools/precision_probe.py \
  --result results_nusc.json \
  --dataroot data/nuscenes \
  --version v1.0-trainval \
  --split val \
  --K 100

python3 tools/precision_probe_p50_p100.py \
  --result results_nusc.json \
  --dataroot data/nuscenes \
  --version v1.0-trainval \
  --split val

Documentation entry points

  • Landing page: /
  • Architecture reference: /architecture/
  • BEVFormer comparison: /bevformer-comparison/
  • Experiment tables: /experiments/
  • Code mapping: /api-reference/

Notes

Validation status. The command examples above reflect the repository surface but were not executed end-to-end in the current documentation environment because the MMDetection3D and CUDA runtime stack is not installed here.