Lifelong Person Re-identification via Knowledge Refreshing and Consolidation

Lifelong person re-identification (LReID) is in significant demand for real-world development as a large amount of ReID data captured from diverse locations over time and cannot be accessed at once inherently. However, key challenge LReID how to incrementally preserve old knowledge gradually add new capabilities the system. Unlike most existing methods, which mainly focus on dealing with catastrophic forgetting, our more challenging problem, is, not only trying reduce forgetting tasks but also aiming improve model performance both during lifelong learning process. Inspired by biological process human cognition where somatosensory neocortex hippocampus work together memory consolidation, we formulated called Knowledge Refreshing Consolidation (KRC) that achieves positive forward backward transfer. More specifically, refreshing scheme incorporated rehearsal mechanism enable bi-directional transfer introducing dynamic an adaptive working model. Moreover, consolidation operating dual space further improves stability long-term. Extensive evaluations show KRC’s superiority state-of-the-art methods pedestrian benchmarks. Code available https://github.com/cly234/LReID-KRKC.