Self-Constrained Network for Re-indexing Free Lifelong Person Re-identification

Re-indexing Free Lifelong Person Re-Identification (RFL-ReID) is a branch of L-ReID that pays particular attention to data privacy scenarios, where old knowledge cannot be revisited through re-indexing when training new tasks. This poses new challenges for knowledge handling and memory capabilities. To address these challenges, we designed a self-constrained networks(SCN) for re-indexing free lifelong person re-identification that includes a Fisher Information Matrix Constraint module and a Feature Self-Fusion module. These components help enhance the model's resistance to catastrophic forgetting in the absence of re-indexing. Additionally, we extensively explored the dynamic relationship between Fisher information changes, model resistance to catastrophic forgetting, and model complexity in the context of lifelong person re-identification tasks. Experiments on several benchmark Reid datasets demonstrate that our method achieves significant performance improvements in both RFL-ReID and traditional L-ReID tasks.