High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification

Wang, Guan’an; Yang, Shuo; Liu, Huanyu; Wang, Zhicheng; Yang, Yang; Wang, Shuliang; Yu, Gang; Zhou, Erjin; Sun, Jian

doi:10.1109/cvpr42600.2020.00648

Cited by 355 publications

(169 citation statements)

References 23 publications

Supporting

Mentioning

169

Contrasting

Order By: Relevance

“…On the CUHK03 dataset, we compare 12 methods with our solution, and the results are shown in Table 3. These various state-of-the-art methods include local feature models [7,57,60,63], attention mechanism models [35,36,55], multibranch networks [8,25,30], multiscale networks [56,62,64], and other deep models [45,49,53,54,58,61]. It T A B L E 1 Statistics of Market-1501, DukeMTMC-reID, and CUHK03 datasets, wherein N id , N im , and N cam represent the number of the identity, image, and camera, respectively.…”

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

Multi‐branch network with hierarchical bilinear pooling for person reidentification

et al. 2021

View full text Add to dashboard Cite

Because of issues such as viewpoint changes, posture variations, and background cluttering, the task of person reidentification (Re-ID) remains challenging. The model of combining global features and part features has been widely used in person Re-ID technology in recent years, but these efforts ignored feature interaction between the convolutional layers and thus lost detailed information conducive to identifying pedestrians under different cameras. To achieve interaction between hierarchical features, a multibranch network with hierarchical bilinear pooling (MBN-HBP) is proposed. The network consists of a global branch, a part-level branch, and a hierarchical bilinear pooling (HBP) branch. The person features extracted by the network include not only global and part-level features but also detailed HBP features. The final feature representation will be more robust to deal with the complex surveillance environment. By conducting comprehensive experiments, competitive performance on the Market-1501, DukeMTMC-Re-ID, and CUHK03 datasets is obtained.This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes.

show abstract

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

Multi‐branch network with hierarchical bilinear pooling for person reidentification

et al. 2021

View full text Add to dashboard Cite

show abstract

“…For occlusion problems, Luo et al 29 use a spatial transform module to align the partial ones by transforming the holistic image, and calculate the distance between aligned pairs. Wang et al 30 present a novel framework to solve the occlusion problems by learning high‐order relation. Kniaz et al 31 propose a ThermalGAN framework for cross‐modality color‐thermal person re‐id, which solves the cross‐modal problem by using a stack of generative adversarial networks (GANs) to translate a single color probe image to a multimodal thermal probe set and extracting features in the infrared images (IR) domain.…”

Section: Related Workmentioning

confidence: 99%

Topology and channel affinity reinforced global attention for person re‐identification

Wang

Gao

Xin

et al. 2021

Int J of Intelligent Sys

View full text Add to dashboard Cite

Person re‐identification (re‐id), with the goal to recognize persons from images captured by nonoverlapping cameras, is a challenging topic in computer vision. It has been studied extensively in recent years, and the attention mechanism is widely applied in person re‐id. But many works mainly focus on extracting discriminative features from local saliency regions, while ignoring some potentially global information between whole‐body features and body‐part features. In this study, we first proposed two effective global information for extracting discriminative features: spatial topology information (STI) and channel affinity information (CAI). On this basis, we further propose a Multi‐information Fusion reinforced Global Attention (MIFGA) module which can effectively fuse a variety of information and utilize more comprehensive information to guide the learning of attention, so as to obtain pedestrian features that are conducive to clustering. Specifically, the proposed MIFGA module includes spatial attention (MIFGA‐S) and channel attention (MIFGA‐C). MIFGA‐S mainly utilizes local feature semantic information and STI to guide the learning of spatial attention. Furthermore, to mine the potential topology information in original feature maps, the self‐learning graph convolution network is proposed. MIFGA‐C fuses channel semantic information and CAI to guide the learning of channel attention. Extensive ablation studies demonstrate that our proposed MIFGA significantly enhances the baseline model and achieves a competitive performance compared with the state‐of‐the‐art person re‐id methods on standard data sets Market‐1501, DukeMTMC‐reID, and CUHK03.

show abstract

“…In the context of cross-domain, the method of feature extraction restricts the accuracy of ReID models. Some scholars use CNN to extract pedestrian global features such as color, shape, texture and etc [8], [29]. For instance, both HHL [45] and ECN [46] model use GAN to transfer images from each camera to the others in target domain and utilize CNN to extract pedestrian global features to identify most of distinctive features, but they cannot distinguish between details and thus miss some features.…”

Section: Related Workmentioning

confidence: 99%

“…Although ReID algorithms based on deep learning have achieved remarkable results, it is still a challenge whether models can be efficiently applied to multiple datasets. For a model such as HOReID proposed by Wang et al which has achieved the Rank-1 identification rate of 94.2% on Market1501, when it is trained and tested on two different datasets, both gene ralization and accuracy of models decreased [29]. This cross domain problem is often caused by domain bias such as different lighting [14] and camera angles [21], [49] among several datasets, greatly limiting the generalizability of models to real-life situations.…”

Section: Introductionmentioning

confidence: 99%

“…Furthermore, when we extracting features, pedestrian images are frequently occluded by foreign objects, such as bicycles, stones, handrails, etc [29], [43]. As the automatic detection does not contain complete appearances of pedestrians, it might cause distinct pedestrians with the same occlusion pattern regarded as the same one or semantic analysis errors.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Cross-Domain Person Re-Identification Based on Feature Fusion

Luo

Ouyang²,

et al. 2021

IEEE Access

View full text Add to dashboard Cite

Person re-identification (ReID) is one of the commonly used criminal investigation methods in reconnaissance. Although the current ReID has achieved robust results on single domains, the focus of researches has shifted to cross-domain in recent years, which is caused by domain bias between different datasets. Generative Adversarial Networks (GAN) is used to realize the image style transfer of different datasets to alleviate the effect of cross-domain [9], [13]. However, the existing GAN-based models ignore complete expressions and occlusion of pedestrian characteristics, resulting in low accuracy in feature extraction [12], [22]. To address these issues, we introduce a cross domain model based on feature fusion (FFGAN) to fuse global, local and semantic features to extract more delicate pedestrian features. Before extracting pedestrian features, we preprocess feature maps with a feature erasure block to solve an occlusion issue. Finally, FFGAN enables a more complete visual description of pedestrian characteristics, thereby improving the accuracy of FFGAN in identifying pedestrians. Experimental results show that the effect of FFGAN is significantly improved compared with some advanced cross-domain ReID algorithms.

show abstract

High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification

Cited by 355 publications

References 23 publications

Multi‐branch network with hierarchical bilinear pooling for person reidentification

Multi‐branch network with hierarchical bilinear pooling for person reidentification

Topology and channel affinity reinforced global attention for person re‐identification

Cross-Domain Person Re-Identification Based on Feature Fusion

Contact Info

Product

Resources

About