A Bottom-Up Clustering Approach to Unsupervised Person Re-Identification

Abstract

Most person re-identification (re-ID) approaches are based on supervised learning, which requires intensive manual annotation for training data. However, it is not only resourceintensive to acquire identity annotation but also impractical to label the large-scale real-world data. To relieve this problem, we propose a bottom-up clustering (BUC) approach to jointly optimize a convolutional neural network (CNN) and the relationship among the individual samples. Our algorithm considers two fundamental facts in the re-ID task, i.e., diversity across different identities and similarity within the same identity. Specifically, our algorithm starts with regarding individual sample as a different identity, which maximizes the diversity over each identity. Then it gradually groups similar samples into one identity, which increases the similarity within each identity. We utilizes a diversity regularization term in the bottom-up clustering procedure to balance the data volume of each cluster. Finally, the model achieves an effective trade-off between the diversity and similarity. We conduct extensive experiments on the large-scale image and video re-ID datasets, including Market-1501, DukeMTMCreID, MARS and DukeMTMC-VideoReID. The experimental results demonstrate that our algorithm is not only superior to state-of-the-art unsupervised re-ID approaches, but also performs favorably than competing transfer learning and semi-supervised learning methods.

Unsupervised Person Re-identification via Cross-Camera Similarity Exploration

“…On Market-1501, under the same setting, we obtain the best performance among the compared methods with rank-1 = 71.7%, mAP = 37.8%. Compared to the state-of-the-art unsupervised method BUC [14], we achieve 10.7 points and 7.2 points improvement on rank-1 accuracy and mAP, respectively. On DukeMTMC-reID, compared to BUC, our method achieves 12.3 and 6.7 points of improvement on rank-1 accuracy and mAP, respectively.…”

“…On MARS, we obtain rank-1 = 62.8%, mAP = 43.6%. Compared to BUC [14], We achieve 4.9 and 8.9 points of improvement in rank-1 accuracy and mAP, respectively. On DukeMTMC-VideoReID, we achieve rank-1 of 76.4% and mAP of 69.3%, which beats BUC by 0.2 and 1.0 points, respectively.…”

“…Following [33,32,14], we adopt the classification model with a non-parametric classifier, where a lookup table is used to store the features of all training images. The stored feature of each image is then used as the weight vector of each class.…”

“…As shown in Table 1 Table 3. Comparison with BUC [14] on Market-1501 and DukeMTMC-reID. The column Auxiliary lists the auxiliary information utilized by the method.…”

“…CCE encourages cross-camera image selection, which enables the model learning from diverse images and getting robust to camera views. Besides, we evaluate CCE based on BUC [14], and the result is shown in Table 3. We observe that on Market-1501, the improvements of using CCE are 4.9 points for rank-1 and 1.2 points for mAP.…”