Mitigate Catastrophic Remembering via Continual Knowledge Purification for Noisy Lifelong Person Re-Identification

Kunlun Xu, Haozhuo Zhang, Yu Li, Yuxin Peng, Jiahuan Zhou

Abstract: Current lifelong person re-identification (LReID) methods focus on tackling a clean data stream with correct labels. When noisy data with wrong labels are given, their performance is severely degraded since the model inevitably and continually remembers erroneous knowledge induced by the noises. Moreover, the well-known catastrophic forgetting issue in LReID becomes even more challenging since the correct knowledge contained in the old model is disrupted by noisy labels. Such a practical noisy LReID task is important but challenging, and rare works attempted to handle it so far. In this paper, we initially investigate noisy LReID by proposing a Continual Knowledge Purification (CKP) method to address the catastrophic remembering of erroneous knowledge and catastrophic forgetting of correct knowledge simultaneously. Specifically, a Cluster-aware Data Purification module (CDP) is designed to obtain a cleaner subset of the given noisy data for learning. To achieve this, the label confidence is estimated based on the intra-identity clustering result where the high-confidence data are maintained. Besides, an Iterative Label Rectification (ILR) pipeline is proposed to rectify wrong labels by fusing the prediction and label information throughout the training epochs. Therefore, the noisy data are rectified progressively to facilitate new model learning. To handle the catastrophic remembering and forgetting issues, an Erroneous Knowledge Filtering (EKF) algorithm is proposed to estimate the knowledge correctness of the old model, and a weighted knowledge distillation loss is designed to transfer the correct old knowledge to the new model while excluding the erroneous one. Finally, a Noisy LReID benchmark is constructed for performance evaluation and extensive experimental results demonstrate that our proposed CKP method achieves state-of-the-art performance.

Primary Subject Area: [Engagement] Multimedia Search and Recommendation

Secondary Subject Area: [Experience] Multimedia Applications

Relevance To Conference: This paper investigates the lifelong person re-identification with label noise. It is inherently a multimedia task because, on the one hand, person re-identification aims to achieve cross-camera person matching which involves multiple camera devices at different positions and times. On the other hand, we study the lifelong learning scenario where the cameras from new cities or viewpoints show up continually. However, existing lifelong person re-identification mainly focuses on the cleanly labeled data to investigate the catastrophic forgetting problem. However, we reveal that existing lifelong person re-identification is vulnerable to noisy labels, i.e., the images are assigned the wrong labels, with deteriorated catastrophic forgetting and degraded acquisition capacity. Therefore, we propose a novel noise-robust method to facilitate the model to perform lifelong learning under data with label noise. We believe that label noise frequently occurs in other multimedia and multimodal tasks, and our methods can bring new insights to relevant research.

Submission Number: 341