enhancement(pose): evaluate ROI vs full-frame inference for the active backend

[kmch4n]

Summary

The current pipeline first detects people with YOLOv8, crops a padded ROI, and then runs YOLO11-Pose inside that ROI. This keeps the pipeline modular, but it may reduce pose quality when the crop clips limbs/equipment or removes global context.

Expected behavior

Evaluate whether YOLO11-Pose should run on the full frame, or on a larger/aspect-preserved crop, with track-aware person selection. Keep the current ROI path only if it wins on representative skiing footage.

Current behavior

YOLOv8 supplies bboxes, Yolo11Backend.estimate(...) crops a padded ROI, and YOLO11-Pose selects the highest-confidence person inside that cropped ROI.

Evidence

• README.md:75 documents ROI-based pose estimation.
• src/skisense/main.py:440 runs YOLO person detection.
• src/skisense/main.py:448 iterates detected bboxes.
• src/skisense/backends/yolo11_backend.py:119 to src/skisense/backends/yolo11_backend.py:132 pads and crops the ROI before pose inference.
• src/skisense/backends/yolo11_backend.py:174 to src/skisense/backends/yolo11_backend.py:197 selects the highest-confidence detected person inside the ROI.

Impact

Deep ski angulation often pushes wrists, ankles, skis, and poles near bbox edges. A crop-first pose path can lose context before the pose model runs, which may reduce keypoint stability even with YOLO11-Pose.

Suspected cause

The pipeline reuses a detector-first architecture even though YOLO11-Pose can already detect people and keypoints together.

Priority

medium

Labels

enhancement

Acceptance criteria

• Add an experiment comparing current ROI inference against full-frame or larger context inference on representative skiing frames.
• Use track association, bbox overlap, or target ID to choose the skier when multiple people are present.
• Measure keypoint confidence, score stability, and inference time.
• Document the chosen mode and any config switch if both modes are kept.

Notes

This is not a duplicate of #3. That issue is about adding a different SOTA backend; this issue keeps YOLO11-Pose and evaluates the inference strategy around it.

[kmch4n]

Re-scoped after the SAM 3D Body migration (#11).

The original framing assumed YOLO11-Pose was the primary engine and that no full-frame path existed. Both have changed:

• YOLO11-Pose is no longer the default. SKISENSE_POSE_BACKEND selects the engine; the default is now SAM 3D Body (3D MHR), with YOLO11-Pose as a selectable lightweight backend. The crop-vs-context question now applies primarily to whichever backend is active.
• A full-frame path already exists. --fast runs estimate_full_frame(...) (implemented on both backends), feeding the whole frame instead of a per-detection ROI. Part of the original acceptance criteria is therefore already satisfied.

Updated scope:

• Compare ROI inference vs full-frame (--fast) for the active backend (SAM 3D Body by default) on representative skiing footage.
• Keep track-aware target selection (already implemented via longest mode) in the comparison.
• Measure keypoint stability, score stability, and inference time; document which mode wins and keep the config switch if both are retained.

Note: deep angulation pushing wrists/ankles near bbox edges is now partly mitigated since SAM 3D Body recovers feet and wrists from the MHR mesh.