AGENTS.md

AGENTS.md

Summary

This is the repository reference for person detection plus human pose estimation with Lite-HRNet. Use it when you need body keypoints on detected people rather than a generic detector.

Use This Example When

• You need human pose estimation.
• You want a person-detection first stage with a swappable pose model.
• You need camera or replay input with packaged standalone support.

Do Not Use This Example When

• You need hand or animal pose instead of human pose.
• You need a single-stage detector.
• You need multi-person tracking rather than per-frame pose overlays.

Quick Facts

• Category: neural-networks/pose-estimation/human-pose
• Shape: script+standalone
• Primary task: person detection plus human pose estimation
• Entrypoint: main.py
• Standalone path: oakapp.toml
• Frontend: none
• Runs on: RVC2 peripheral, RVC4 peripheral, and RVC4 standalone packaging
• Requires: person detector and Lite-HRNet-style pose model
• Input: camera frames by default or ReplayVideo via --media_path
• Output: Video, Detections, and Pose
• Models: YOLOv6 and Lite-HRNet YAMLs in depthai_models/
• Visualizer / UI: DepthAI Visualizer via dai.RemoteConnection

Read First

• README.md
• main.py
• utils/annotation_node.py
• utils/arguments.py

Architecture

• A person detector runs first.
• ImgDetectionsFilter identifies the person class for the pose workflow.
• FrameCropper extracts padded person crops for Lite-HRNet.
• GatherData and utils/annotation_node.py merge the keypoints and skeleton back to the original detections.

Constraints

• The current code path is person-specific, even though the detector could emit other classes.
• The pose parser threshold is intentionally set to 0.0 so the overlay node can do the filtering instead.
• Replay sizing and crop padding affect downstream pose quality.

Related Examples

• neural-networks/pose-estimation/hand-pose: use this when you need hand landmarks and gesture logic
• neural-networks/pose-estimation/animal-pose: use this when you need animal pose
• neural-networks/reidentification/human-reidentification: use this when you need to identify tracked people or faces rather than estimate pose

Validation

• Run: python3 main.py
• Success looks like: the Visualizer shows Video, Detections, and Pose, and visible people receive skeleton overlays
• Common failure meaning: the detector is not stably finding people, crop generation drifted, or the selected pose model does not match the parser assumptions