ExtendedKeypointMetrics and choose between depthai and luxonis-train post-processing in the keypoint parser

[dtronmans]

Purpose

• Can pass sigmas and area_factor into ExtendedKeypointMetrics for the metrics calculation to match luxonis-train config
• The keypoint parser now accepts postprocessing: luxonis_train (default is depthai, which keeps the current behavior). depthai: normalized keypoints, luxonis_train: pixel-space
• Arbitrary number of keypoints per class and arbitrary number of classes

Change with output names (this is an optional/separate change): just match the ONNX outputs by name (like kpt_output* like in depthai-nodes) instead of relying on positional ordering

Example parser in the config:

parser:
  name: YOLOKeypointDetectionParser
  params:
    subtype: yolov8
    n_classes: 3
    conf_thres: 0.03
    iou_thres: 0.65
    max_det: 300
    postprocessing: luxonis_train

Example keypoint metric:

metrics:
  metrics:
    - name: ExtendedKeypointMetrics
      params:
        kpt_oks_sigmas: [0.08, 0.08, 0.08, 0.08]
        area_factor: 1.0

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation