Estimate Pose (Ultralytics YOLO)

How it fits

Typical backends

Caveats

• Hard constraintThe bound weights directory is scanned recursively and must contain exactly one weights file; zero aborts startup with a no-weights error and two or more with an ambiguous-binding error. The checkpoint head is read and must be a pose head — a detection, segmentation, OBB, or classification checkpoint aborts at startup.
• I/O contractThe output shape is fixed at deploy time by the connected consumer, and all arms run identical inference so the keypoint set agrees across them; the same keypoints are flattened, grouped per person, paired with the person box, or wrapped in skeleton topology.
• Hard constraintThe [([Landmark], BoundingBox)] arm requires the checkpoint to emit a person box, and the [Skeleton] arm carries the fixed COCO-17 edge set so requires exactly 17 keypoints; a checkpoint that violates either aborts that frame.
• AccuracyEach Landmark confidence is the per-keypoint visibility score; a non-default checkpoint may change the keypoint count or order, so the keypoint catalogue must be set to match the loaded weights.
• Parameter interactionThe square inference size should match the checkpoint's training input size (640 for the default), or accuracy silently degrades; the confidence and IoU thresholds gate person detection and its non-maximum suppression.
• FallbackA device beginning with cuda silently falls back to CPU with a warning when CUDA is unavailable.
• State lifetimeAll configuration is captured once at startup; runtime changes have no effect and require a redeploy.