Abstract
Lifelong object re-identification incrementally learns from a stream of re-identification tasks. The objective is to learn a representation that can be applied to all tasks and that generalizes to previously unseen re-identification tasks. The main challenge is that at inference time the representation must generalize to previously unseen identities. To address this problem, we apply continual meta metric learning to lifelong object re-identification. To prevent forgetting of previous tasks, we use knowledge distillation and explore the roles of positive and negative pairs. Based on our observation that the distillation and metric losses are antagonistic, we propose to remove positive pairs from distillation to robustify model updates. Our method, called Distillation without Positive Pairs (DwoPP), is evaluated on extensive intra-domain experiments on person and vehicle re-identification datasets, as well as inter-domain experiments on the LReID benchmark. Our experiments demonstrate that DwoPP significantly outperforms the state-of-the-art.
Introduction
Object Re-Identification (ReID) is a critical computer vision task focused on matching the same individual or vehicle across non-overlapping camera views. In real-world deployments, ReID systems must operate in a "lifelong" manner, continuously adapting to new environments and camera setups. This presents two major hurdles: catastrophic forgetting, where learning new data erodes performance on previous tasks, and the requirement for zero-shot generalization, as the system must identify individuals it has never encountered during training.
While many continual learning methods exist for image classification, ReID is fundamentally a metric learning problem. The goal is to learn an embedding space where features of the same identity are clustered together (positive pairs) and features of different identities are pushed apart (negative pairs). Furthermore, privacy regulations often prohibit the storage of "exemplars" (raw images from previous tasks), necessitating exemplar-free learning strategies.
Continual Meta Metric Learning
To address these challenges, we propose a framework based on Continual Meta Metric Learning. We build upon Deep Meta Metric Learning (DMML), which frames ReID as a series of few-shot learning episodes. This meta-learning approach encourages the model to learn a feature extractor that generalizes better to unseen identities compared to standard global optimization techniques.
Distillation without Positive Pairs (DwoPP)
A common strategy to mitigate forgetting in exemplar-free settings is Knowledge Distillation (KD). KD encourages the current model to produce probability distributions similar to those of the previous model. However, we identify a fundamental conflict between the metric loss and the distillation loss regarding "positive pairs" (the ground-truth class of a query image).
In a new task, the old model may not be confident about a new identity, resulting in a "flat" probability distribution for that class. Standard KD (Distillation with Positive Pairs, or DwPP) tries to force the new model to match this low-confidence output. Simultaneously, the metric learning loss is trying to maximize the confidence for the positive class to create tight clusters. This antagonism destabilizes training.
Our solution, Distillation without Positive Pairs (DwoPP), removes the positive class from the distillation process. By focusing distillation only on the relationships between negative classes ("dark knowledge"), we allow the metric loss to optimize the positive pairs without interference, while the distillation loss preserves the structure of the rest of the feature space.
Experimental Results
We evaluated DwoPP on several large-scale benchmarks, including Market-1501 and MSMT17_V2 for person ReID, and VeRi-776 for vehicle ReID. We also tested inter-domain generalization on the LReID benchmark.
Our results show that DwoPP consistently outperforms existing state-of-the-art (SOTA) methods across all datasets. For instance, on Market-1501, DwoPP achieves a final mAP of 67.2% after 10 tasks, significantly higher than the 52.5% achieved by the previous SOTA method, AKA. Notably, DwoPP even outperforms some replay-based methods that are allowed to store image exemplars, highlighting the effectiveness of our refined distillation strategy.
Conclusion
This work demonstrates that standard knowledge distillation can be counterproductive in lifelong metric learning scenarios. By identifying and removing the antagonistic influence of positive pairs in the distillation loss, DwoPP provides a robust, exemplar-free solution for lifelong object re-identification. Our approach sets a new state-of-the-art, proving that careful consideration of loss interactions is key to successful incremental learning in embedding-based tasks.
Summary updated