---

name: inference description: "Run and adapt MMDetection3D point-cloud, monocular, multi-modality, and lidar segmentation inference workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MMDetection3D inference sub-skill

Use this sub-skill when the user needs to run, modify, or troubleshoot MMDetection3D inference for:

• LiDAR / point-cloud 3D detection.
• Monocular image 3D detection.
• LiDAR + image multi-modality 3D detection.
• LiDAR point-cloud semantic segmentation.
• Python API calls through mmdet3d.apis or the 3D inferencer classes.

This sub-skill is intentionally safe by default: the bundled script only renders commands and never loads models, checkpoints, datasets, GPUs, or visualizers.

Start here

1. Identify the task type:
   • lidar-det: point-cloud detection.
   • mono-det: monocular image detection with an annotation/info file.
   • multi-modality-det: point cloud plus image(s) with an annotation/info file.
   • lidar-seg: point-cloud segmentation.
2. Confirm the user has a matching config and checkpoint for the same model family, dataset, classes, and task.
3. For demo-style command construction, use scripts/build_inference_command.py. It prints a shell command only.
4. For Python integration, use references/api-reference.md to choose between low-level API functions and inferencer classes.
5. For output, visualization, remote-server, and batch workflow choices, use references/workflows.md.
6. For failures, use references/troubleshooting.md.

Route away

• Dataset conversion, dataset root layout, or annotation/info generation belongs in the data-preparation sub-skill.
• Config-family selection, model zoo alias lookup, or config mutation belongs in the configuration-model-zoo sub-skill.
• Training, testing, metric evaluation, TTA, or distributed launch belongs in the training-evaluation sub-skill.
• Geometry objects, coordinate conversions, box projection, or visualization internals belong in the structures-visualization sub-skill.
• TorchServe, model publishing, FLOPs, or log-analysis utilities belong in the serving-tools sub-skill.

Minimal operating rules

• Do not run an inference command unless the user explicitly asks for model execution and accepts the hardware/checkpoint/runtime cost.
• Prefer no-display-safe commands on remote servers: omit --show, keep an --out-dir, and save predictions.
• Demo commands are file-path workflows. For in-memory arrays, use the inferencer classes where supported; see the ndarray limitations in references/api-reference.md.
• Treat checkpoint URLs and model aliases as potentially network-triggering during actual execution. The bundled command builder does not download anything.
• Always preserve the user's config/checkpoint pairing unless they ask for model-family changes.