SVDNet for Pedestrian Retrieval

By Yifan Sun, Liang Zheng, Weijian Deng, Shengjin Wang. IEEE International Conference on Computer Vision, 2017

798 citations recorded by the source corpus.

This paper proposes the SVDNet for retrieval problems, with focus on the application of person re-identification (reID). We view each weight vector within a fully connected (FC) layer in a convolutional neuron network (CNN) as a projection basis. It is observed that the weight vectors are usually highly correlated. This problem leads to correlations among entries of the FC descriptor, and compromises the retrieval performance based on the Euclidean distance. To address the problem, this paper proposes to optimize the deep representation learning process with Singular Vector Decomposition (SVD). Specifically, with the restraint and relaxation iteration (RRI) training scheme, we are able to iteratively integrate the orthogonality constraint in CNN training, yielding the so-called SVDNet. We conduct experiments on the Market-1501, CUHK03, and DukeMTMC-reID datasets, and show that RRI effectively reduces the correlation among the projection vectors, produces more discriminative FC descriptors, and significantly improves the re-ID accuracy. On the Market-1501 dataset, for instance, rank-1 accuracy is improved from 55.3% to 80.5% for CaffeNet, and from 73.8% to 82.3% for ResNet-50.

Subfields: Computer Vision and Pattern Recognition, Ocean Engineering

View Paper: https://www.semanticscholar.org/paper/f45c568e40e08ee1d75e4f063db3a96a392c0032

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. Luc Van Gool
3. Dacheng Tao
4. Ming-Hsuan Yang
5. Xiaogang Wang
6. Ling Shao
7. Tao A. Xiang
8. Qi Tian

In Ocean Engineering

1. Yi Yang
2. Liang Zheng
3. Shaogang Gong
4. Xiatian Zhu
5. Tao A. Xiang
6. Wanli Ouyang
7. Mang Ye
8. Zhun Zhong

Citations in the corpus

• AANet: Attribute Attention Network for Person Re-Identifications (2019, neutral): “5 SVDNet [21] 56.”
• Deep Cosine Metric Learning for Person Re-identification (2018, neutral): “40 SVDNet [23]a 82.”
• Interaction-And-Aggregation Network for Person Re-Identification (2019, methodology): “The training procedure of IANet follows the standard identity classification paradigm [39, 53, 59], where the identify of each person is treated as a distinct class.”
• Dual Attention Matching Network for Context-Aware Feature Sequence Based Person Re-identification (2018, methodology): “, [10, 4, 19, 37, 41, 17, 12, 45, 35]; and feature set or feature sequence based methods, e.”
• Batch Coherence-Driven Network for Part-Aware Person Re-Identification (2020, neutral): “4 SVDNet [75] 76.”
• Camera Style Adaptation for Person Re-identification (2017, methodology): “Compared with published IDE implementations [29, 47, 43], IDE∗ has the best rank-1 accuracy on Market-1501.”
• Adaptively Connected Neural Networks (2019, methodology): “+Re-ranking [55], PAN, DPFL [4], and the newly proposed methods SVDNet [39], TriNet + Era.”
• Self-Referenced Deep Learning (2018, methodology): “Following [41,30], we train the network by the cross-entropy loss (Eq (2)) and use the feature layer’s output as the representation of person bounding box images for test by the Euclidean distance metric.”
• Pose Transferrable Person Re-identification (2018, neutral): “0 SVDNet [30] 40.”
• Person Re-identification by Deep Learning Multi-scale Representations (2017, background): “2 As this dataset was newly constructed for person re-id from the multitarget multi-camera tracking benchmark DukeMTMC [46], there are only a small number of results reported in a few unpublished arXiv papers [78, 31, 52].”
• Mask-Guided Contrastive Attention Model for Person Re-identification (2018, methodology): “The compared methods include the Deep Pyramid Feature Learning (DPFL) [10], SVDNet [33], and two re-ranking methods: DaF [42] and Re-ranking [52].”
• Beyond Part Models: Person Retrieval with Refined Part Pooling (2017, methodology): “We note that the IDE model specified in [40] is a commonly used baseline in deep re-ID systems [40, 38, 33, 10, 28, 42, 43, 45].”
• Part-Aligned Bilinear Representations for Person Re-identification (2018, neutral): “1 5 10 20 mAP 1 5 10 20 mAP Varior et al. 2016 [52] 61:6 - - - 35:3 - - - - - Zhong et al. 2017 [71] 77:1 - - - 63:6 - - - - - Zhao et al. 2017 [18] 80:9 91:7 94:7 96:6 63:4 - - - - - Sun et al. 2017 [73] 82:3 92:3 95:2 - 62:1 - - - - - Geng et al. 2016 [74] 83:7 - - - 65:5 89:6 - - - 73:8 Lin et al”
• Harmonious Attention Network for Person Re-identification (2018, methodology): “Importantly, the performance advantage by our method is achieved at a [31] 65.8 39.5 76.0 48.4 HPN [22] 76.9 - - - SVDNet [28] 82.”
• In Defense of the Triplet Loss for Person Re-Identification (2017, background): “son re-identification (ReID) has attracted significant attention in the computer vision community. Especially with the rise of deep learning, many new approaches have been proposed to achieve this task [40,8, 42,31,39,10,52,4,46,20,54,35]1. In many aspects person ReID is similar to image retrieval, wh”
• Human Semantic Parsing for Person Re-identification (2018, neutral): “9 SVDNet [36] 56.”
• Multi-level Factorisation Net for Person Re-identification (2018, methodology): “Deep Neural Networks for Person Re-ID Most recent person Re-ID methods train deep DNN models with various learning objectives including classification, verification and triplet ranking losses [42, 38, 24, 32, 27].”