Introduction
Person Re-identification (ReID) is a critical computer vision task aimed at matching individuals across non-overlapping camera views. While standard ReID assumes a static dataset, Lifelong ReID (LReID) addresses the more realistic scenario where a model must learn from a continuous stream of data. However, a significant yet often overlooked challenge in LReID is the "clothing change" problem. Over long periods, individuals inevitably change their outfits, rendering clothing-based features—which are highly discriminative in the short term—unreliable.
This paper introduces the Cloth-Hybrid Lifelong Person Re-identification (CH-LReID) setting. Unlike traditional LReID, which assumes cloth consistency, CH-LReID involves a data stream containing both cloth-consistent and cloth-changing samples. To address this, the authors propose the Differentiated Knowledge Consolidation (DKC) framework.
The DKC Framework
The DKC framework is designed to balance the retention of old knowledge with the acquisition of new, cloth-robust features. It consists of three primary components:
1. Differentiated Knowledge Transfer (DKT)
Standard transfer learning often treats all incoming data uniformly. DKT, however, employs unsupervised clustering (DBSCAN) to identify fine-grained appearance patterns. By calculating an adaptive coefficient, the model determines the reliability of clothing features for a given sample. If a person's outfit has changed significantly relative to previous stages, DKT prioritizes learning cloth-irrelevant identity features to ensure robust matching.
2. Latent Knowledge Consolidation (LKC)
To mitigate "catastrophic forgetting"—the tendency of neural networks to lose previously learned information when trained on new data—DKC introduces LKC. Instead of simple output distillation, LKC focuses on latent space reconstruction. It utilizes a transfer network to map current features back into the feature space of previous models. This ensures that the structural relationships between identities remain stable even as the backbone network evolves to accommodate new data.
3. Dual-level Distribution Alignment (DDA)
DDA addresses the inherent conflict between cloth-relevant and cloth-irrelevant information. It aligns feature distributions at two distinct granularities:
- Instance Level: Ensuring consistency for individual identities.
- Fine-grained Level: Aligning clusters of similar appearances to maintain a nuanced understanding of identity that isn't solely dependent on broad clothing categories.
Experimental Results
The DKC framework was evaluated against a suite of seven ReID datasets, including both cloth-consistent (e.g., Market-1501, MSMT17) and cloth-changing (LTCC, PRCC) varieties.
Performance Metrics
DKC demonstrated superior performance across multiple training sequences. On a representative "Order-1" sequence, the model achieved:
- 44.2% mAP and 55.8% Rank-1 accuracy, significantly outperforming previous state-of-the-art methods like DKP and PatchKD.
- Average Forgetting (AF) of 3.9%, a marked improvement over the 6.2% seen in competing methods, highlighting the effectiveness of the LKC module.
Qualitative Analysis
Visualizations using t-SNE confirmed that DKC produces more compact and well-separated clusters for different identities. Crucially, these clusters remain distinct even when the same individual appears in different outfits, a scenario where traditional models typically fail by grouping samples based on clothing similarity rather than true identity.
Conclusion
The DKC framework represents a significant step forward in making Person ReID systems viable for long-term real-world deployment. By explicitly modeling the "Cloth-Hybrid" nature of lifelong data streams and employing differentiated transfer and latent space reconstruction, DKC effectively balances the need for discriminative features with the necessity of cloth-invariance. Future work may focus on reducing the training complexity of the multi-module pipeline and improving robustness to extreme domain shifts, such as day-to-night transitions.
Summary updated