Bi-C2R: Bidirectional Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification

TL;DR

This paper tackles privacy-aware lifelong person re-identification by introducing Re-indexing Free Lifelong Re-ID ($RFL\text{-}ReID$) and a Bidirectional Continual Compatible Representation (Bi-C2R) framework. Bi-C2R combines a Bidirectional Compatible Transfer Network (BiCT-Net) with Bidirectional Compatible Distillation (BiCD), Bidirectional Anti-forgetting Distillation (BiAD), and a Dynamic Feature Fusion (DFF) module to update historical gallery features without re-indexing, while balancing old and new knowledge across diverse domains. Theoretical analysis and extensive experiments on seven datasets and multiple training orders show state-of-the-art performance for RFL-ReID and strong anti-forgetting behavior on traditional L-ReID tasks, along with significant inference-time efficiency gains. The approach enables privacy-preserving, scalable lifelong Re-ID with robust cross-domain feature compatibility and minimal degradation of old knowledge, making it practically impactful for real-world surveillance systems.

Abstract

Lifelong person Re-IDentification (L-ReID) exploits sequentially collected data to continuously train and update a ReID model, focusing on the overall performance of all data. Its main challenge is to avoid the catastrophic forgetting problem of old knowledge while training on new data. Existing L-ReID methods typically re-extract new features for all historical gallery images for inference after each update, known as "re-indexing". However, historical gallery data typically suffers from direct saving due to the data privacy issue and the high re-indexing costs for large-scale gallery images. As a result, it inevitably leads to incompatible retrieval between query features extracted by the updated model and gallery features extracted by those before the update, greatly impairing the re-identification performance. To tackle the above issue, this paper focuses on a new task called Re-index Free Lifelong person Re-IDentification (RFL-ReID), which requires performing lifelong person re-identification without re-indexing historical gallery images. Therefore, RFL-ReID is more challenging than L-ReID, requiring continuous learning and balancing new and old knowledge in diverse streaming data, and making the features output by the new and old models compatible with each other. To this end, we propose a Bidirectional Continuous Compatible Representation (Bi-C2R) framework to continuously update the gallery features extracted by the old model to perform efficient L-ReID in a compatible manner. We verify our proposed Bi-C2R method through theoretical analysis and extensive experiments on multiple benchmarks, which demonstrate that the proposed method can achieve leading performance on both the introduced RFL-ReID task and the traditional L-ReID task.

Figures (11)

• Figure 1: Comparison between the Traditional Lifelong person Re-identification (L-ReID) task (a) and our introduced Re-indexing Free Lifelong person Re-identification (RFL-ReID) task (b).
• Figure 2: Performance Comparison of existing methods, including LwF li2017learning, AKA pu2021lifelong, CVS wan2022continual, PatchKD sun2022patch, DKP xu2024distribution.
• Figure 3: A schematic diagram of our proposed Bidirectional Continual Compatible Representation (Bi-C2R) method.
• Figure 4: The architecture of our proposed Bidirectional Continual Compatible Representation (Bi-C2R) method. Our Bi-C2R consists of a Bidirectional (Forward/Backward) Compatible Transfer network (BiCT-Net), a Bidirectional Compatible Distillation (BiCD) module, and a Bidirectional Anti-forgetting Distillation (BiAD) module. In the training phase, all the above components will be trained, and the BiCT-Net (forward) will be used to update the old feature set. In the updating phase, a Dynamic Feature Fusion (DFF) module is employed to update the old gallery features.
• Figure 5: The structure of our proposed Bidirectional Compatible Transfer network (BiCT-Net).