PCBEAR: Pose Concept Bottleneck for Explainable Action Recognition

Human pose sequences offer a more interpretable alternative to black-box AI models for action recognition: PCBEAR uses skeletal body movements rather than pixel-level features or text descriptions, making it easier for humans to understand why the system classified a particular action the way it did.

The framework distinguishes between two types of motion explanations: Static pose concepts capture body positions in individual frames, while dynamic pose concepts track motion patterns across multiple frames—this dual approach better captures the temporal dependencies that are essential for understanding human actions.

Concepts can be automatically discovered without manual labeling: By applying clustering to video pose sequences, PCBEAR identifies meaningful motion patterns on its own, reducing the need for human experts to manually define what constitutes an "explainable concept."

The method maintains strong performance while improving transparency: Testing on three datasets (KTH, Penn-Action, and HAA500) demonstrated that PCBEAR achieves high classification accuracy while simultaneously providing interpretable, motion-driven explanations—showing that explainability doesn't require sacrificing accuracy.

This approach addresses a critical gap in video AI transparency: Existing explainability methods struggle with motion dynamics and temporal patterns in videos; PCBEAR's focus on pose-based concepts directly tackles this limitation, which is important for real-world applications requiring accountability.