---

name: structures-visualization description: "Work with MMDetection3D box, point, coordinate-conversion, and visualization geometry." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MMDetection3D structures and visualization sub-skill

Use this sub-skill for:

• 3D box origin, yaw, and mode questions.
• LiDAR / Camera / Depth box and point conversions.
• points_cam2img / points_img2cam projection checks.
• Det3DLocalVisualizer and saved geometry artifacts from demos or evaluation.

Start here

1. Identify the coordinate family and whether the object is a box, point cloud, or image projection artifact.
2. Use references/geometry-api.md for origin, yaw, mode, and projection rules.
3. Use references/visualization.md for visualizer methods and saved-output conventions.
4. Run scripts/inspect_geometry.py for a tiny synthetic smoke check when the installed package is available.
5. Use references/troubleshooting.md when a box looks shifted, mirrored, clipped, or empty.

Route away

• End-to-end inference and prediction outputs belong in inference.
• Dataset browsing or conversion belongs in data-preparation.
• Training, testing, TTA, or distributed launch belongs in training-evaluation.
• Config choice or adaptation belongs in configuration-model-zoo.

Operating rules

• Treat box origin as part of the data contract.
• Keep LiDAR, Camera, and Depth modes explicit in every conversion.
• Prefer bundled references and script output over source-repo paths.
• Do not rely on GUI display during the smoke check.