Informative Attention Based Deep Network For Person Re-Identification

Abstract

Person re-identification (ReID) is the most essential task in visual based surveillance. The recent success of deep learning in extracting unique and powerful discriminatory features has made immense shift in this era. The existing limitations in major challenges of person re-identification like pose variations, body misalignment, patch based learning are not comprehensively addressed. We propose Informative Attention Based IAB Network, a new framework that integrate multiple attention modules by preserving localized and global contextual information so that model can learn most informative, representative and discriminative features. A deep convolutional neural network is used to train the base model which is then used as a feature extractor. Since Res Net is well known in re-id task, so the backbone network of IAB Net is based on Res Net architecture and after adding attention mechanism in Res Net we named it Attention Based ResNet-50. Specifically, we have also introduced Channel and Spatial Attention (CASA) Network that consists on a pair of attention modules named as Channel Attention Module and Spatial Attention Module. Channel attention module and spatial attention module primarily focusing on channel aggregation, spatial dimension and position awareness, respectively. We have added COAB on the output of each convolutional block of ResNet-50. After generating informative feature map from base network, IAB-Net split into a global network and CASA network in parallel. In our proposed pipeline, we have used this pair after each convolutional block of ResNet-50, that significantly boost the performance and representation power of network. By using this new lightweight backbone with orthogonality constraint to enforce diversity on both hidden activation and weights and along with attention modules, our extensive experiments on different popular Re-ID datasets i.e DukeMTMC-reID and Market-1501 have achieved state-of-the-art performance and we confirm that our model manifest harmonious refinement in detection and classification. Explicitly, on Market-1501 it achieved 89.88% Rank-1 and 75.85% mAP. On Duke MTMC-ReID IAB-Net achieved 79.20% Rank-1 and 64.04% mAP and on PRID-450s we achieved 37.33% Rank-1, 74.89% Rank-10 and 83.78% Rank-20, which fundamentally perform well against all existing approach. Our experiment and results endorse that IAB-Net significantly perform well on both performance measure rank-1 and mAP.