Lifelong Person Re-identification by Pseudo Task Knowledge Preservation

Abstract

In real world, training data for person re-identification (Re-ID) is collected discretely with spatial and temporal variations, which requires a model to incrementally learn new knowledge without forgetting old knowledge. This problem is called lifelong person re-identification (LReID). Variations of illumination and background for images of each task exhibit task-specific image style and lead to task-wise domain gap. In addition to missing data from the old tasks, task-wise domain gap is a key factor for catastrophic forgetting in LReID, which is ignored in existing approaches for LReID. The model tends to learn task-specific knowledge with task-wise domain gap, which results in stability and plasticity dilemma. To overcome this problem, we cast LReID as a domain adaptation problem and propose a pseudo task knowledge preservation framework to alleviate the domain gap. Our framework is based on a pseudo task transformation module which maps the features of the new task into the feature space of the old tasks to complement the limited saved exemplars of the old tasks. With extra transformed features in the task-specific feature space, we propose a task-specific domain consistency loss to implicitly alleviate the task-wise domain gap for learning task-shared knowledge instead of task-specific one. Furthermore, to guide knowledge preservation with the feature distributions of the old tasks, we propose to preserve knowledge on extra pseudo tasks which jointly distills knowledge and discriminates identity, in order to achieve a better trade-off between stability and plasticity for lifelong learning with task-wise domain gap. Extensive experiments demonstrate the superiority of our method as compared with the state-of-the-art lifelong learning and LReID methods.

LSTKC: Long Short-Term Knowledge Consolidation for Lifelong Person Re-identification

“…To overcome this problem, various methods have been proposed which can be mainly categorized into two branches: data replay-based methods and knowledge distillation-based ones. The data reply-based approaches aim to prevent knowledge forgetting via storing and replaying exemplars from historical datasets (Wu and Gong 2021;Ge et al 2022;Yu et al 2023;Chen, Lagadec, and Bremond 2022;Huang et al 2022). However, such a strategy tends to hinder data privacy and incur substantial computational overheads.…”

“…The knowledge distillation technique is widely used in LReID (Pu et al 2021;Wu and Gong 2021;Ge et al 2022;Sun and Mu 2022) by forcing the new model to generate consistent outputs as the old model. (Pu et al 2021) was one of the initial works that introduced Knowledge distillation into the LReID task and adopted logistic distillation which forced the new model to generate the same classification score as the old model.…”

“…(Pu et al 2022) claimed the crucial role of Batch Normalization (BN) in data distribution shifting and proposed reconciliation normalization to constrain the learning of BN layers. Apart from the aforementioned exemplar-free LReID methods, various exemplar-based approaches also adopt knowledge distillation to alleviate catastrophic forgetting (Wu and Gong 2021;Ge et al 2022;Yu et al 2023).…”

“…where γ is a hyperparameter to balance two components. Following the existing works (Ge et al 2022), for the first training step, the parameters of Θ 1 are initialized with the ImageNet pre-trained model, and the parameters of Φ 1 are randomly initialized. For the later steps, the parameters of Θ t are initialized with Θ t−1 and the parameters of Φ t−1 which is a linear layer are initialized with the mean feature of the identities in D t .…”

“…Similar to other lifelong learning tasks (Liang et al 2022), the challenge of catastrophic forgetting emerges as a critical obstacle due to the discrepancy in knowledge across diverse datasets. To handle this issue, several LReID approaches aim to retain exemplars from the old datasets as the rehearsal of historical knowledge for the learning of new models (Wu and Gong 2021;Ge et al 2022;Yu et al 2023). However, this solution will indeed impede data privacy and suffer from considerable computational overheads.…”