---

name: bev-former description: "Routes BEVFormer camera-only 3D detection workflows, from install/import and config inspection through nuScenes data prep, distributed training and evaluation, and log analysis." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

BEVFormer

Use this skill for the BEVFormer repository when the task mentions BEVFormer, BEVFormerV2, nuScenes camera-only 3D detection, BEV perception, or the projects.mmdet3d_plugin OpenMMLab plugin.

Start here

• Read model zoo and config map when you need a family overview or a config knob refresher.
• Run the bundled environment checker for a quick import/config/CUDA smoke from any working directory.
• Read repository provenance if you need to know whether this skill matches the current checkout before using or refreshing it.
• Read troubleshooting when setup, import, data, checkpoint, or runtime errors appear.

Route map

installation-and-configs

Use this route for:

• legacy OpenMMLab install and import checks
• plugin wiring, config inheritance, and config summaries
• BEVFormer vs BEVFormerV2 architecture questions
• model-family selection and BEV/temporal knob questions

Read sub-skills/installation-and-configs/SKILL.md and run its bundled config inspector when you need a static summary.

dataset-preparation

Use this route for:

• nuScenes raw tree validation
• CAN-bus expansion placement
• temporal nuscenes_infos_temporal_*.pkl generation or validation
• data_root / ann_file layout questions

Read sub-skills/dataset-preparation/SKILL.md and run its bundled layout checker when you need missing-path diagnostics.

training-and-evaluation

Use this route for:

• distributed training commands
• distributed evaluation commands
• FP16 command composition
• checkpoint, launcher, and work-dir questions
• warnings about --eval, --format-only, or multi-GPU eval behavior

Read sub-skills/training-and-evaluation/SKILL.md and use the command builders instead of hand-editing shell launchers.

analysis-and-utilities

Use this route for:

• JSON or JSONL log summaries
• routing benchmark or visualization requests
• checkpoint utility caveats
• safe analysis helpers that do not train or mutate checkpoints

Read sub-skills/analysis-and-utilities/SKILL.md and use the bundled log summarizer for small fixtures.

Public prerequisites

• Python 3.8 is the documented baseline.
• The legacy stack documented by the repo uses torch 1.9.1+cu111, mmcv-full 1.4.0, mmdet 2.14.0, mmsegmentation 0.14.1, and mmdet3d 0.17.1.
• BEVFormer data workflows assume nuScenes plus the CAN-bus expansion.
• Training, evaluation, and visualization are checkpoint- and GPU-dependent; use the command builders and references first.

Bundled helpers

• scripts/check_bevformer_environment.py — run this first when imports or config parsing look broken.
• sub-skills/installation-and-configs/scripts/inspect_bevformer_config.py — use for a deeper static config summary.
• sub-skills/dataset-preparation/scripts/check_bevformer_nuscenes_layout.py — use for missing-data-path diagnostics.
• sub-skills/training-and-evaluation/scripts/bevformer_train_command.py and sub-skills/training-and-evaluation/scripts/bevformer_eval_command.py — use for copyable train/eval commands.
• sub-skills/analysis-and-utilities/scripts/summarize_bevformer_log.py — use for tiny log summaries.

Minimal smoke check

If you have a checkout on disk, pass it explicitly to the helper so the script does not depend on shell activation state:

python scripts/check_bevformer_environment.py --repo-root <checkout-root> --config projects/configs/bevformer/bevformer_tiny.py

For a V2 config, point --config at projects/configs/bevformerv2/bevformerv2-r50-t1-base-24ep.py.

Before refresh

Compare the current checkout against repository provenance before deciding whether to refresh this skill.