Inter-Intra Cluster Reorganization for Unsupervised Vehicle Re-Identification

State-of-the-art unsupervised object re-identification (Re-ID) methods conduct model training with pseudo labels generated by clustering techniques. Unfortunately, due to the existence of inter-ID similarity and intra-ID variance problems in vehicle Re-ID, clustering sometimes mixes different similar vehicles together or splits images of the same vehicle in different views into different clusters. To enhance the model’s ID discrimination capability in the presence of such kinds of label noise, we propose an inter-intra cluster reorganization approach (ICR) to reorganize the relationship between instances within and between clusters, which can provide higher-quality contrastive learning guidance based on existing clustering results. In the intra-cluster reorganization, we design a camera-aware maximum reliability sub-cluster organization approach, which reorganizes each cluster into several intersecting sub-clusters of higher quality based on the finer intra-camera clustering results. We further design a novel metric called centroid reliability to measure the reliability of intra-cluster contrastive learning. In the inter-cluster reorganization, we propose an ambiguous cluster discrimination criterion to measure the probability that two clusters belong to the same vehicle. Based on this criterion, we design a focal contrastive loss to adaptively re-organize the contribution of ambiguous clusters in model training to perform better contrastive learning. Extensive experiments on VeRi-776 and VERI-Wild demonstrate that ICR is effective and can achieve state-of-the-art performance.

I. Introduction

Vehicle re-identification (Re-ID) aims to identify the same vehicle captured by different cameras. It can be used for vehicle cross-camera tracking and vehicle trajectory mining, and has great potential value in intelligent transportation and public security management due to the massive deployment of surveillance cameras on the city road. Previous research [1], [2], [3], [4], [5] focus on supervised vehicle Re-ID, which involves model training and testing in annotated datasets from the same domain, and have achieved good performance. Unfortunately, due to the domain gap between the training datasets and actual application scenarios [6], [7], [8], [9], models trained on public datasets cannot be directly applied in practice since the performance will be greatly degraded. A large amount of application scenario data must be collected and annotated with vehicle identities to obtain a practical model, which is impractical. On the one hand, since most surveillance videos are not clear enough to recognize the plate number, it’s time-consuming to label vehicle identity based on appearance alone. On the other hand, this approach does not enable the effective utilization of the massive data generated by surveillance cameras every day. Therefore, the unsupervised Re-ID method that can learn discriminative representations without annotation is more practical and has recently attracted increasing attention.