Disentangling Static and Dynamic Information for Reducing Static Bias in Action Recognition

2025-09-26

AI Summary

Action recognition models often suffer from “static bias,” wherein they over-rely on static scene cues (e.g., background, objects), leading to poor generalization—especially in zero-shot settings. To address this, we propose a dual-stream disentanglement framework that explicitly separates static (biased) and dynamic (unbiased) representations. We enforce statistical independence between the two streams via an independence loss and further constrain the static stream to encode only scene information using a scene prediction loss, thereby suppressing its interference with action classification. The method requires no additional annotations and is plug-and-play compatible with mainstream architectures. Experiments across multiple benchmarks demonstrate substantial mitigation of static bias: on zero-shot action recognition, our approach achieves an average accuracy improvement of 8.2%. Moreover, it enhances robustness in real-world scenarios and improves model interpretability.

Technology Category

• Computer Vision: Scene Analysis & Understanding
• Machine Learning: Calibration & Uncertainty Quantification
• Knowledge Representation and Reasoning: Action, Change, and Causality

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Action recognition models rely excessively on static cues rather than dynamic human motion, which is known as static bias. This bias leads to poor performance in real-world applications and zero-shot action recognition. In this paper, we propose a method to reduce static bias by separating temporal dynamic information from static scene information. Our approach uses a statistical independence loss between biased and unbiased streams, combined with a scene prediction loss. Our experiments demonstrate that this method effectively reduces static bias and confirm the importance of scene prediction loss.

Problem

• Reducing static bias in action recognition models
• Separating temporal dynamic from static scene information
• Improving performance in real-world and zero-shot scenarios

Innovation

• Separates temporal dynamic from static scene information
• Uses statistical independence loss between biased streams
• Combines with scene prediction loss to reduce bias