Impact Statement:Person re-identification (ReID) focuses on building the correspondence between human appearances across deployed non-overlapping cameras. As an essential issue in compute...Show More
Abstract:
Person re-identification (ReID) aims to identify pedestrian images with the same identity across non-overlapping camera views. Intra-camera supervised person re-identific...Show MoreMetadata
Impact Statement:
Person re-identification (ReID) focuses on building the correspondence between human appearances across deployed non-overlapping cameras. As an essential issue in computer vision, person ReID can be applied to intelligent video surveillance, security system, and other fields. However, the difficulty in obtaining inter-camera identity labels hinders the real-world application of supervised ReID methods. Unsupervised ReID approaches dispense with the expensive annotation, but their performance needs improvement. This paper proposes a plug-and-play strategy for ICS-ReID, a semi-supervised ReID paradigm that only leverages intra-camera identity labels to address the issue. This strategy consists of two independent and complementary modules to incorporate the intra-camera supervisory information into the unsupervised ReID framework, thus significantly reducing the workload of annotating identities across cameras and achieving promising results for ReID.
Abstract:
Person re-identification (ReID) aims to identify pedestrian images with the same identity across non-overlapping camera views. Intra-camera supervised person re-identification (ICS-ReID) is a new paradigm that trains a model using only intra-camera labels, thus reducing the cost of inter-camera identity association. Pseudo-label-based clustering algorithms perform well in the unsupervised ReID task, whereas they inevitably generate noisy pseudo labels through clustering, especially in the early training stage. Given this, we propose an unsupervised pseudo-labeling method to help in the semi-supervised ICS-ReID task. This method improves the clustering results by reassigning pseudo labels for the training data and consists of two modules, Absorb and Repel. The Absorb module aims to group all data with the same intra-camera identity into one cluster. The Repel module ensures that images under the same camera view but with different identities do not appear in the same cluster. Both modul...
Published in: IEEE Transactions on Artificial Intelligence ( Volume: 5, Issue: 6, June 2024)