Re-ranking Person Re-identification with k-Reciprocal Encoding

By Zhun Zhong, Liang Zheng, Donglin Cao, Shaozi Li. Computer Vision and Pattern Recognition, 2017

1,649 citations recorded by the source corpus.

When considering person re-identification (re-ID) as a retrieval process, re-ranking is a critical step to improve its accuracy. Yet in the re-ID community, limited effort has been devoted to re-ranking, especially those fully automatic, unsupervised solutions. In this paper, we propose a k-reciprocal encoding method to re-rank the re-ID results. Our hypothesis is that if a gallery image is similar to the probe in the k-reciprocal nearest neighbors, it is more likely to be a true match. Specifically, given an image, a k-reciprocal feature is calculated by encoding its k-reciprocal nearest neighbors into a single vector, which is used for re-ranking under the Jaccard distance. The final distance is computed as the combination of the original distance and the Jaccard distance. Our re-ranking method does not require any human interaction or any labeled data, so it is applicable to large-scale datasets. Experiments on the large-scale Market-1501, CUHK03, MARS, and PRW datasets confirm the effectiveness of our method.

Subfields: Computer Vision and Pattern Recognition, Ocean Engineering

View Paper: https://www.semanticscholar.org/paper/76d2422cf5db5a6a96977fc42dcdd92422c27c21

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Yi Yang
2. Dacheng Tao
3. Luc Van Gool
4. Xiaogang Wang
5. Ming-Hsuan Yang
6. Ling Shao
7. Liang Zheng
8. Tao A. Xiang

In Ocean Engineering

1. Yi Yang
2. Liang Zheng
3. Tao A. Xiang
4. Shaogang Gong
5. Xiatian Zhu
6. Dapeng Chen
7. Wanli Ouyang
8. Hongsheng Li

Citations in the corpus

• Deep Learning for Person Re-Identification: A Survey and Outlook (2020, methodology): “It is now widely used in Re-ID community [33], [49], [68], [75], [76].”
• Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification (2020, methodology): “Randomly erasing (Zhong et al., 2017b) is only adopted in target-domain fine-tuning.”
• Camera Style Adaptation for Person Re-identification (2017, neutral): “00 Re-rank [48] 77.”
• SVDNet for Pedestrian Retrieval (2017, methodology): “The re-ranking method [36] is higher than ours in mAP on Market-1501, because re-ranking exploits the relationship among the gallery images and results in a high recall.”
• Bilinear Attention Networks for Person Retrieval (2019, methodology): “We adopt the new training/testing protocol proposed in [53].”
• Unsupervised Person Re-Identification via Multi-Label Classification (2020, methodology): “Label filtering by cycle consistency: Inspired by kreciprocal nearest neighbor[12, 44], we assume that, if two images belong to the same class, their neighbor image sets should also be similar.”
• Reinforced Temporal Attention and Split-Rate Transfer for Depth-Based Person Re-identification (2017, background): “As such, there is vast literature in learning discriminative distance metrics [35,42,105,51,55,77,43,61,53,19,48,108,91,5] and discriminant subspaces [63,47,107,84,43,46,64,94,15].”
• In Defense of the Triplet Loss for Person Re-Identification (2017, result): “In the following section we evaluate them experimentally and show that, for ReID, they achieve superior perfor- mance compared to both the traditional triplet loss and the previously published variants of it [7, 25].”
• Furnishing Your Room by What You See: An End-to-End Furniture Set Retrieval Framework with Rich Annotated Benchmark Dataset (2019, methodology): “methods are considered to refine the retrieval results like [34] [3], while in our method a context embedding model is employed to collaboratively improve the search results of a furniture set.”
• Pose Transferrable Person Re-identification (2018, methodology): “For CUHK03, we train the ResNet-50 with cross-entropy loss and triplet loss respectively, i.e., the DenseNet-169 model is not used because its performance is similar to the ResNet-50 model on this dataset.”
• Generalizing a Person Retrieval Model Hetero- and Homogeneously (2018, methodology): “The IDE-based methods [39,40,41] achieve good performance on fully labeled datasets, but often fail to generalize to a new target set.”
• The Devil is in the Details: Self-Supervised Attention for Vehicle Re-Identification (2020, methodology): “For the case of VeRi and VERI-Wild datasets, we also try the method of re-ranking suggested in [49] and achieved considerable mAP scores of 82.”
• Exploit the Unknown Gradually: One-Shot Video-Based Person Re-identification by Stepwise Learning (2018, methodology): “Following recent works [7, 17, 38], we take the training process as an identity classification task.”
• Attention-Aware Compositional Network for Person Re-identification (2018, background): “For Market-1501, CUHK03-NP and DukeMTMC-reID, rank-1 identification rate and mean Average Precision (mAP) are reported.”