Notes on Human Pose Estimation Pipelines

Pose estimation quality is often discussed as a model-selection problem. In practice, the surrounding pipeline can dominate the result.

Capture first

Motion blur, rolling shutter, occlusion, and a moving camera create errors that downstream models cannot fully repair. A stable viewpoint and short exposure can matter more than replacing one strong pose model with another.

Track one subject well

For rowing analysis, a clean single-athlete crop is more valuable than unstable multi-person detections. Tracking errors become discontinuities in joint trajectories and then appear again as false velocities and accelerations.

Lift with context

Monocular 3D lifting is ambiguous frame by frame. Temporal context helps the model preserve plausible motion, but camera geometry and unusual rowing positions can still produce systematic depth errors.

Treat smoothing as a modeling choice

Smoothing should remove detector noise without erasing the rapid changes that define catch and finish timing. The filter belongs inside the evaluation, not as an invisible preprocessing detail.