---

name: data-preparation description: "Prepare MMDetection3D datasets, custom dataset layouts, conversion commands, and info-file migrations." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Data Preparation

Use this sub-skill when the task is to prepare, inspect, or troubleshoot dataset roots and conversion commands for MMDetection3D v1.4.0.

Use this for

• Validating dataset root layouts for KITTI, Waymo, NuScenes, Lyft, SemanticKITTI, S3DIS, ScanNet, SUN RGB-D, and custom datasets.
• Rendering safe dataset-conversion commands for the repository's native data converter without executing downloads or conversions.
• Planning v1 info-file migrations and coordinate-update commands before touching user data.
• Explaining ready-made annotation files, ground-truth databases, and dataset-specific preparation caveats.

First steps

1. Identify the dataset family, dataset root, desired stage (source, preconvert, or converted), version, and whether the user only needs command construction or also layout validation.

2. Read references/datasets.md for built-in dataset layouts and conversion command families.

3. For custom data, read references/custom-dataset.md before suggesting a converter or config.

4. Run the bundled safe checker when paths are available:

   python scripts/check_dataset_layout.py kitti --root data/kitti --stage preconvert
   

5. Render, but do not execute, conversion or migration commands with:

   python scripts/build_create_data_command.py create-data kitti --root-path data/kitti --out-dir data/kitti --extra-tag kitti
   

6. Read references/troubleshooting.md when conversion hangs, files are missing, info pickles are stale, or custom data fails to load.

Route away

• Model/config inheritance, class count changes, and model-zoo selection: use configuration-model-zoo.
• Training, testing, distributed launch, metric submission commands, and evaluator prefix flags: use training-evaluation.
• Dataset visualization, coordinate-system debugging, box projections, or geometry APIs: use structures-visualization.
• Implementing a new dataset class, transform, sampler, or registry component: use customization-extensions after using this sub-skill for the data schema.
• Inference input formatting after the dataset is already prepared: use inference.

Safety boundaries

The bundled scripts in scripts/ are intentionally non-mutating: they only validate local path existence or print commands. They never download datasets, convert files, update pickles, launch training, or import large SDKs. Before running any rendered command, copy or back up valuable annotation files and confirm that the command is being run from a MMDetection3D checkout that contains the native conversion tools.