Scalable memory synthesis for exemplar-free lifelong person re-identification.

Abstract

Lifelong person re-identification (LReID) aims to learn from streaming data sources step by step, which suffers from catastrophic forgetting. In the exemplar-free LReID setting, some existing works generate pseudo-images to approximate past-domain data distributions in order to alleviate forgetting. However, such pixel-level replay strategies mainly align low-level appearance statistics such as color and style, and often neglect explicit constraints on the geometric structure of old identities in the embedding space, which tends to lead to continuous performance degradation and aggravated forgetting on past domains. To address this issue, this paper proposes Scalable Memory Synthesis (SMSyn), which constructs a controllable synthetic memory to maintain instance-level separability during continual updates. The Prototype Variational Synthesizer (PVS) generates pseudo-samples around historical instances to enrich and reconstruct historical knowledge. Then, the Feature Space Preservation (FSP) module constrains samples to concentrate around their corresponding class centers while maintaining inter-class relationships and local neighborhood structures, thereby counteracting geometric inflation and suppressing the accumulation of identity-geometry degradation. Experiments demonstrate that SMSyn outperforms state-of-the-art LReID methods.