Grounding Foundational Vision Models with 3D Human Poses for Robust Action Recognition

TL;DR

This paper introduces a novel model that combines world dynamics and explicit human pose data to improve action recognition in complex, occlusive scenes, emphasizing the importance of spatial understanding.

Contribution

The paper presents a new architecture that fuses predictive world dynamics with explicit human pose data for more robust action recognition.

Findings

Outperforms baseline models on InHARD and UCF-19-Y-OCC benchmarks.

Excels in complex scenes with occlusions.

Highlights the importance of spatial grounding over pattern recognition.

Abstract

For embodied agents to effectively understand and interact within the world around them, they require a nuanced comprehension of human actions grounded in physical space. Current action recognition models, often relying on RGB video, learn superficial correlations between patterns and action labels, so they struggle to capture underlying physical interaction dynamics and human poses in complex scenes. We propose a model architecture that grounds action recognition in physical space by fusing two powerful, complementary representations: V-JEPA 2's contextual, predictive world dynamics and CoMotion's explicit, occlusion-tolerant human pose data. Our model is validated on both the InHARD and UCF-19-Y-OCC benchmarks for general action recognition and high-occlusion action recognition, respectively. Our model outperforms three other baselines, especially within complex, occlusive scenes. Our findings emphasize a need for action recognition to be supported by spatial understanding instead of statistical pattern recognition.