Improving Person Re-identification by Mask Guiding and Part Pooling
Bin Zhang, Yanfeng Li, Houjin Chen, Jia Sun · 2020
Person re-identification (re-ID) is a promising computer vision task. State-of-the-art methods mainly utilize deep learning based approaches to learn visual features for describing person appearances. Due to occlusion, complex background, different postures and light intensity, the technology faces many challenges. In this paper, a person re-identification method is proposed combining mask guiding and part pooling. First, person mask is generated by a segmentation sub-net, combining Macro-Micro Adversarial Network (MMAN) with Fully Convolution Networks (FCN). To alleviate the influence of background, a mask guiding strategy is designed integrating the mask with person feature map. Then a part pooling strategy is employed to extract local features of the person. The final loss function of the network is defined as the combination of global loss and local loss, which can describe the person in a comprehensive manner. Four public datasets are employed to test the performance of the proposed method. Our method achieves rank-1/mAP of 89.05%/72.83% on the Market-1501, 79.03%/62.06% on the DukeMTMC-reID, 46.79%/29.61% on the MSMT-17, 49.50%/45.03% on the CUHK03-NP. Experimental results show that the designed mask guiding and part pooling strategies can improve person re-ID performance.