Topology-guided transfer learning for cross-domain human action recognition

Abstract

Achieving robust human action recognition (HAR) under cross-domain and low-resource conditions remains challenging due to substantial distribution shifts and limited annotated data. Existing transfer learning methods primarily rely on static knowledge transfer or feature distribution alignment, limiting their ability to capture higher-order structural relationships and adaptively transfer knowledge across domains. To address these limitations, we propose a topology-guided transfer learning framework that leverages source-domain structural knowledge to facilitate target-domain adaptation. Specifically, we construct an unsupervised topology-preserving graph to capture the intrinsic structure of the source-domain feature space, which serves as a transferable topological memory. A gated attention fusion module dynamically retrieves and integrates the structural information most relevant to each target-domain sample. Finally, topology-aware alignment and knowledge distillation are jointly employed to preserve structural consistency during training and distill topological knowledge into the target model for memory-free inference. We evaluate the proposed method on two cross-domain transfer tasks using UCF101 as the shared source domain, with HMDB51 and a self-collected dataset serving as the respective target domains. Experimental results demonstrate that our approach consistently outperforms representative transfer learning baselines across multiple backbones and cross-domain settings. These findings confirm that integrating transferable topological memory with adaptive knowledge transfer effectively improves cross-domain HAR performance.