AlphaPose-GCN / Pr-VIPE (Chen et al.)

Highlights

• Incorporates element priors into the graph convolution to disambiguate quarter-turn differences.
• Pr-VIPE variant adds probabilistic embeddings to capture annotator disagreement in GOE labels.
• Supports real-time inference on consumer GPUs for rink-side analytics.

Implementation Notes

AlphaPose confidence scores act as attention weights inside the GCN; ensure thresholds mirror the training regime. The authors suggest temperature scaling when transferring to community datasets like YourSkatingCoach.