Cross-View Asymmetric Metric Learning for Unsupervised Person Re-Identification

By Hong-Xing Yu, Ancong Wu, Weishi Zheng. 2017

341 citations recorded by the source corpus.

While metric learning is important for Person reidentification (RE-ID), a significant problem in visual surveillance for cross-view pedestrian matching, existing metric models for RE-ID are mostly based on supervised learning that requires quantities of labeled samples in all pairs of camera views for training. However, this limits their scalabilities to realistic applications, in which a large amount of data over multiple disjoint camera views is available but not labelled. To overcome the problem, we propose unsupervised asymmetric metric learning for unsupervised RE-ID. Our model aims to learn an asymmetric metric, i.e., specific projection for each view, based on asymmetric clustering on cross-view person images. Our model finds a shared space where view-specific bias is alleviated and thus better matching performance can be achieved. Extensive experiments have been conducted on a baseline and five large-scale RE-ID datasets to demonstrate the effectiveness of the proposed model. Through the comparison, we show that our model works much more suitable for unsupervised RE-ID compared to classical unsupervised metric learning models. We also compare with existing unsupervised REID methods, and our model outperforms them with notable margins. Specifically, we report the results on large-scale unlabelled RE-ID dataset, which is important but unfortunately less concerned in literatures.

Subfields: Biomedical Engineering, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

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

In Biomedical Engineering

1. Liang Zheng
2. Yi Yang
3. Shaogang Gong
4. Qi Tian
5. Xiatian Zhu
6. Shiliang Zhang
7. Mang Ye
8. Tao A. Xiang

In Computer Vision and Pattern Recognition

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

In Representation & Self-Supervised Learning

1. Shaogang Gong
2. Liang Zheng
3. Yi Yang
4. Tao A. Xiang
5. Yanwei Fu
6. Michael Tschannen
7. Sylvain Gelly
8. J. Lai

Citations in the corpus

• On Mutual Information Maximization for Representation Learning (2019, methodology): “When using different encoders for different views and asymmetric critics as employed by CPC, DeepInfoMax, and CMC one recovers asymmetric variants of (5), see, e.g. (Yu et al., 2017; Zhang et al., 2019).”
• Deep Spatial Feature Reconstruction for Partial Person Re-identification: Alignment-free Approach (2018, methodology): “Besides, several state-of-the-art methods are compared, including Bag of Words (BOW) [31], Multi-scale Contextaware Network (MSCAN) [11], Spindle Net [29], Reranking [29], Consistent-Aware Deep Learning (CADL) [14], Cross-View Asymmetric Metric Learning (CAMEL) [26], DNSL+OL-MANS [35] and Deeply-Lea”
• Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-Identification (2019, methodology): “Then, these methods learn a metric [36] or refine the re-ID model by unsupervised clustering [9] on the target domain.”
• Generalising without Forgetting for Lifelong Person Re-Identification (2021, methodology): “Existing Re-ID methods (Xiao et al. 2016; Yu, Wu, and Zheng 2017; Song et al. 2019) mostly assume that all training data can be prepared in advance for model learning.”
• Self-Similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-Identification (2018, methodology): “Bag-of-Words (BoW) [45] and local maximal occurrence (LOMO) [24], three unsupervised methods, including UMDL [29],PUL [10] and CAMEL [43], and six unsupervised domain adaptation methods, including PTGAN [40], SPGAN [8], TJ-AIDL [39], ARN [23], UDAP [33] and MAR [44].”
• Adaptive Transfer Network for Cross-Domain Person Re-Identification (2019, methodology): “In the UDA community, most of the previous methods [25, 6, 5, 27, 28, 37, 32] attempt to align the source domain to the target domain by reducing the divergence of feature distributions.”
• Interaction-And-Aggregation Network for Person Re-Identification (2019, neutral): “Person reID methods focus on two key points: learning a powerful feature representation for images [8, 43, 14, 44, 24, 26, 28, 42, 1] and designing an effective distance metric [32, 29, 56, 2, 57].”
• Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-identification (2018, background): “However, these methods have a number of limitations that restrict their generalisation: (1) Relying on hand-crafted features without the deep learning capability of automatically learning stronger representations from the training data [36]; (2) Using a pre-learned deep model on labelled source data”
• Deep Learning for Person Re-Identification: A Survey and Outlook (2020, methodology): “The Rank-1 accuracy/mAP increases from 54.5%/26.3% (CAMEL [206]) to 80.0%/58.3% (SSG [213]) on the Market-1501 dataset within two years.”
• FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification (2018, methodology): “Two main types of methods were adopted in previous works, i.e. learning discriminative person representations [30, 31, 32] and metric learning [25, 33, 34, 35].”
• 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].”
• Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification (2017, methodology): “We also compare existing unsupervised methods, including the Clustering-based Asymmetric MEtric Learning (CAMEL) [49], the Progressive Unsupervised Learning (PUL) [6], and UMDL [35].”
• Pose-Normalized Image Generation for Person Re-identification (2017, methodology): “Deep re-id models Most recently proposed re-id models employ a DNN to learn discriminative view-invariant features [25, 2, 9, 51, 34, 40, 63, 47].”
• Deep Representation Learning With Part Loss for Person Re-Identification (2017, background): “Existing approaches conquer this challenge by either seeking discriminative metrics [49, 21, 59, 24, 28, 3, 31, 46, 25, 59, 5, 32, 50], or generating discriminative features [30, 47, 10, 7, 26, 53, 57, 20, 61].”
• Robust Anchor Embedding for Unsupervised Video Person re-IDentification in the Wild (2018, methodology): “[49] introduced a cross-view asymmetric metric learning method to learn the distance measures.”
• Generalizing a Person Retrieval Model Hetero- and Homogeneously (2018, methodology): “We compare with two hand-crafted features, i.e. BoW [44] and LOMO [22], three unsupervised methods, including CAMEL [50], PUL [2], and UMDL [32], and three unsupervised domain adaptation approaches, including PTGAN [6], SPGAN [1] and TJ-AIDL [7].”
• Distilled Person Re-Identification: Towards a More Scalable System (2019, methodology): “We compared with unsupervised ReID methods including unsupervised features LOMO [31], BOW [66] and unsupervised learning models UMDL [42], PTGAN [53], PUL [14], CAMEL [62], SPGAN [13], TJAIDL [51] and HHL [71].”
• Asymmetric metric learning for knowledge transfer (2020, methodology): “In metric learning, asymmetry has been used in sample weighting [40], different mappings per view [80], or hard example mining [78].”