Unsupervised Gait Recognition With Selective Fusion

Abstract

Previous gait recognition methods primarily relied on labeled datasets, which require a labor-intensive labeling process. To eliminate this dependency, we focus on a new task: Unsupervised Gait Recognition (UGR). We introduce a cluster-based baseline to solve UGR. However, we identify additional challenges in this task. First, sequences of the same person in different clothes tend to cluster separately due to significant appearance changes. Second, sequences captured from 0° and 180° views lack distinct walking postures and do not cluster with sequences from other views. To address these challenges, we propose a Selective Fusion method, consisting of Selective Cluster Fusion (SCF) and Selective Sample Fusion (SSF). SCF merges clusters of the same person wearing different clothes by updating the cluster-level memory bank using a multi-cluster update strategy. SSF gradually merges sequences taken from front/back views using curriculum learning. Extensive experiments demonstrate the effectiveness of our method in improving rank-1 accuracy under different clothing and view conditions.