Reproduce “GaitAdapt: Continual Learning for Evolving Gait Recognition” on OpenResearch

About this paper

Jingjie Wang, Shunli Zhang, Xiang Wei, Senmao Tian

The paper introduces GaitAdapt, a new task and evaluation framework for continual gait recognition, and proposes GaitAdapter, a non-replay continual learning method. GaitAdapter utilizes a GaitPartition Adaptive Knowledge (GPAK) module and an Euclidean Distance Stability Method (EDSN) to mitigate catastrophic forgetting and enhance generalization, achieving superior accuracy on various datasets and protocols compared to existing baselines.

Key insights

• The fine-grained nature of gait patterns and the dependency on metric learning necessitate specialized continual learning approaches that stabilize the feature space rather than solely relying on classification output regularization.
• Decomposing gait features into parts and leveraging a graph-based mechanism for knowledge accumulation and transfer effectively manages evolving gait knowledge while preserving old information in a non-replay setting.
• Maintaining the relative Euclidean distances between negative sample pairs is critical for mitigating catastrophic forgetting in metric-learning-based continual gait recognition tasks.

Summary by alphaXiv.