---

name: det3d description: "Guide Det3D PyTorch 3D object-detection workflows across configuration, KITTI/nuScenes/Lyft data preparation, GPU training and evaluation, custom CUDA operations, and visualization." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Det3D

Det3D is a legacy PyTorch toolbox for point-cloud 3D object detection. Use this skill when a task names Det3D, det3d, PointPillars, SECOND, VoxelNet, CBGS, KITTI/nuScenes/Lyft configs, point-cloud voxelization, or its tools/train.py and tools/test.py workflows.

First route

1. Identify the target dataset/version, model family, config, checkpoint, and whether the request is inspection, data preparation, training, evaluation, or visualization.
2. Read compatibility.md and troubleshooting.md before installing or changing dependencies.
3. Run the read-only runtime diagnostic when imports or GPU behavior are uncertain.
4. Choose one focused route:

Request signal	Read
Python configs, registries, PointPillars/SECOND/VoxelNet, anchors, model components	configuration-and-models
KITTI, nuScenes, Lyft roots, info/db files, sweeps, transforms, voxel/data validation	datasets-and-preprocessing
train/test CLI, checkpoints, resume, metrics, distributed/NCCL launch	training-and-evaluation
install/import, CUDA/C++ extensions, spconv, ABI/toolkit, backend diagnostics	runtime-ops
BEV/LiDAR/KITTI display, logs, FLOPs, headless rendering, result inspection	visualization-and-analysis

Cross-route tasks should keep one owner: configuration owns model structure; data owns source schemas and conversion; training owns job/result lifecycle; runtime owns environment/backend diagnosis; visualization owns display/artifact interpretation.

Public prerequisites

The source project documents Linux, Python 3.6+, PyTorch 1.1–1.6, CUDA 10.0/10.1, CMake >=3.13.2, a compatible historical spconv, and dataset SDKs. Treat these as versioned legacy constraints. A modern PyTorch/CUDA installation may import some utilities while remaining incompatible with the custom extensions or sparse-convolution model path. Det3D training/evaluation is documented as GPU-only; CPU import is not a CPU training fallback.

Install in an isolated environment, select a torch/CUDA/toolkit/spconv combination as a unit, then build extensions against that exact environment. Install only the dataset and visualization extras needed by the selected route. Do not download datasets/checkpoints or launch training implicitly.

Minimal checks

python -c "import det3d; print(det3d.__version__)"
python sub-skills/runtime-ops/scripts/check_runtime.py
python sub-skills/datasets-and-preprocessing/scripts/validate_dataset_layout.py DATA_ROOT --dataset kitti
python sub-skills/training-and-evaluation/scripts/plan_launch.py test CONFIG --checkpoint CHECKPOINT

The last two helpers are non-mutating; the dataset checker returns nonzero when required directories are absent, and the launch planner never starts a job. For config-only inspection use sub-skills/configuration-and-models/scripts/inspect_config.py.

Boundaries and safety

• Use user-provided paths for data, configs, checkpoints, and work directories.
• Treat create_data, training, evaluation, distributed launch, extension builds, and interactive visualization as side-effectful or hardware-bound.
• Do not claim an old model-zoo score is reproduced without matching dataset split, preprocessing, evaluator, checkpoint, and dependency versions.
• Do not point future agents to the original checkout's scripts or notebooks; the linked references and helpers are the self-contained replacements.

Read repo-provenance.md when deciding whether this graph is stale relative to a different Det3D checkout.