Eliminating Background-Bias for Robust Person Re-identification Maoqing Tian, Shuai Yi, Hongsheng Li, Shihua Li, Xuesen Zhang, Jianping Shi, Junjie Yan, Xiaogang Wang CVPR 2018 Poster 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Xu Gao, Peking University. gaoxu1024@pku.edu.cn Background Existing models are biased to capture too much relevance between background appearances. 已有方法忽视了background-bias，并且数据集也没有涵盖全，拍摄的camera比较少。 Probe Image Same Person Same Camera Diff Person Same Camera Similar Background Same Person Diff Camera 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Investigation on Background-Bias Build four types of datasets from existing reid datasets. Network framework: Learning Deep Feature Representations with Domain Guided Dropout for Person Re-Identification. CVPR 2016. Foreground Mask: DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. ArXiv 2016. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Investigation on Background-Bias Existing model fails due to background bias. Train on the original dataset. Test on CUHK03. Reason: Overfitting the training dataset with background bias. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Investigation on Background-Bias Background can help distinguish persons? Train on the background-only dataset. Test on CUHK03. The background appearances of the same persons are usually similar in existing datasets, and the networks are easily biased by such similar backgrounds. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Investigation on Background-Bias How to eliminate background-bias? Train on the mean-background and the random-background datasets. Test on CUHK03. Conclusion: Trained model from random-background data should be much more robust to new scenes. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Xu Gao, Peking University. gaoxu1024@pku.edu.cn More Results Test on CUHK03 Test on Market1501 在此处键入公式。 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

The Proposed Framework Similar structure as the whole-person main network, followed by a upsampling module. Output: 4-class parsing map (background + 3 foregrounds) 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

The Proposed Framework Some results of the person parsing network. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Xu Gao, Peking University. gaoxu1024@pku.edu.cn Training Scheme Stage 1: Pretrain the person parsing network with LIP, Human-Parsing and MS-COCO human parsing datasets. Stage 2: The whole-person main network is train independently. Stage 3: Fix parameters of the main network, and train the person- region guided pooling sub-network. Stage 4: The main network and the guided-pooling sub-network are trained end-to-end. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Xu Gao, Peking University. gaoxu1024@pku.edu.cn Experiments Training on HumanParsing, LIP, MS-COCO. Testing on CUHK03, CUHK01, VIPER, 3dpes, Market-1501. Top-K metric. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Xu Gao, Peking University. gaoxu1024@pku.edu.cn Experiments 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Evaluation on Background Influence Dataset Compare with DGD. DGD suffers great performance drop when testing on the mean- background and the random- background dataset. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Analysis of Each Component To Do: More branches … 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn

Xu Gao, Peking University. gaoxu1024@pku.edu.cn Conclusion +Investigation on background-bias. +New ReID network with segmenting person into three parts. -More experiments could be conducted. -The framework is similar with previous works in attention. 2019/3/11 Xu Gao, Peking University. gaoxu1024@pku.edu.cn