Discrepant and Multi-instance Proxies for Unsupervised Person Re-identification

In this supplementary material, we demonstrate more experimental results and implementation details of our method. A. More Experimental Results A.1. Parameter Analysis The cluster contrastive loss weight λ. We analyze the weight λ of cluster contrastive loss L DCP (Eq. 8) in the overall loss L DCM IP (Eq. 10) on Market-1501 and MSMT17 in Figure 1 . λ controls the proportion of the cluster contrastive loss and instance contrastive loss. When λ is small and the instance contrastive loss is weighted more heavily, the performance on both datasets drops significantly, especially for MSMT17. However, when λ is large and the cluster contrastive loss is weighted more heavily, the model still achieves good performance. This suggests that, even after the inclusion of instance-level contrastive learning, cluster-level contrastive learning still contributes more to performance, and that the inter-instance relationships learned based on multi-instance proxies are complementary to the inter-class relationships learned based on discrepant cluster proxies. We set λ = 0.5 because the model achieves the best performance on both datasets at that value. The distance threshold for DBSCAN clustering. In DB-SCAN [3], the clustering threshold is the maximum distance that two samples can have from one another and still be considered neighbors. A larger distance threshold may result in samples with the same ground truth being incorrectly merged, while a smaller distance threshold may result in incorrect splits. Figure 2 shows the sensitivity of our DCMIP to the distance threshold. We find that the smaller threshold is more suitable for the relatively small dataset Market-1501, while the larger threshold is more suitable for the relatively large MSMT17. The optimum distance threshold on Market-1501 and MSMT17 is 0.45 and 0.7, respectively. Although different methods may have differ-