Automated Classification of Baseball Pitching Phases
This paper segments pitching motion into phases using pose estimation and biomechanical features. We draw inspiration from this method to extract keyframes at the point of max knee lift from only extracted poses in a simple and scalable manner.
Interpretable Pre-Release Baseball Pitch Type Anticipation
This work uses 2D to 3D pose estimation and biomechanical feature extraction at key pitching events to classify pitch types with interpretable features. It is relevant because it shows that joint poses and biomechanical signals can reveal meaningful pitch cues. However, its applicability to our use case (batter-perspective, pre-release) is limited by its reliance on events during release.
Spatio-temporal Relation Modeling for Few-shot Action Recognition
This paper introduces a spatio-temporal model combining CNN features, local patch enrichment, and transformer-based temporal reasoning to learn from limited data. Their method addresses domain gaps and small datasets while leveraging full video context. This could help us capture more general, non-pose pitch tipping cues that might be missed by keyframe-only approaches.