DASK: Distribution Rehearsing via Adaptive Style Kernel Learning for Exemplar-Free Lifelong Person Re-Identification

TL;DR

DASK addresses catastrophic forgetting in Lifelong Person Re-Identification by modeling and rehearsing old-domain distributions without retaining historical exemplars. It introduces DRL to learn instance-adaptive distribution transfer via AKPNet and DRRT to consolidate new-old knowledge through joint training on real and generated old-style data, guided by SKD and ReID losses with EMA fusion. The approach yields state-of-the-art performance among exemplar-free LReID methods, with strong anti-forgetting and generalization to unseen domains, validated on comprehensive LReID benchmarks. By enabling privacy-preserving domain rehearsal and robust knowledge consolidation, DASK offers a practical path for scalable lifelong re-identification in dynamic, multi-domain environments.

Abstract

Lifelong person re-identification (LReID) is an important but challenging task that suffers from catastrophic forgetting due to significant domain gaps between training steps. Existing LReID approaches typically rely on data replay and knowledge distillation to mitigate this issue. However, data replay methods compromise data privacy by storing historical exemplars, while knowledge distillation methods suffer from limited performance due to the cumulative forgetting of undistilled knowledge. To overcome these challenges, we propose a novel paradigm that models and rehearses the distribution of the old domains to enhance knowledge consolidation during the new data learning, possessing a strong anti-forgetting capacity without storing any exemplars. Specifically, we introduce an exemplar-free LReID method called Distribution Rehearsing via Adaptive Style Kernel Learning (DASK). DASK includes a Distribution Rehearser Learning (DRL) mechanism that learns to transform arbitrary distribution data into the current data style at each learning step. To enhance the style transfer capacity of DRL, an Adaptive Kernel Prediction Network (AKPNet) is explored to achieve an instance-specific distribution adjustment. Additionally, we design a Distribution Rehearsing-driven LReID Training (DRRT) module, which rehearses old distribution based on the new data via the old AKPNet model, achieving effective new-old knowledge accumulation under a joint knowledge consolidation scheme. Experimental results show our DASK outperforms the existing methods by 3.6%-6.8% and 4.5%-6.5% on anti-forgetting and generalization capacity, respectively. Our code is available at https://github.com/zhoujiahuan1991/AAAI2025-LReID-DASK

Figures (11)

• Figure 1: (a) Cross-domain distribution gap entails knowledge difference and catastrophic forgetting. (b) Knowledge distillation suffers inadequate anti-forgetting capacity. (c) Replay-based methods require an exemplar buffer, compromising data privacy. (d) Our method rehearses old distribution via old-style data generation, possessing a strong anti-forgetting capacity without storing historical exemplars.
• Figure 2: The overall components of our DASK method. At each training step, a new domain dataset $D_t$ is given. (a) DRRT scheme generates the old-style data to enhance knowledge consolidation as the LReID model $\boldsymbol{\mathrm{M}}_t$ learns from $D_t$. (b) DRL mechanism trains an Adaptive Kernel Prediction Network (AKPNet) to achieve instance-specific distribution adjustment, aiming to transform the data of arbitrary domains to domain $t$, preparing for $D_t$-style data generation in subsequent training steps.
• Figure 3: Anti-forgetting tendency on seen domains.
• Figure 4: Generalization tendency on unseen domains.
• Figure 5: Visualization of distribution rehearsing effects.