ABD-Net: Attentive but Diverse Person Re-Identification

Abstract

Attention mechanisms have been found effective for person re-identification (Re-ID). However, the learned "attentive" features are often not naturally uncorrelated or "diverse", which compromises the retrieval performance based on the Euclidean distance. We advocate the complementary powers of attention and diversity for Re-ID, by proposing an Attentive but Diverse Network (ABD-Net). ABD-Net seamlessly integrates attention modules and diversity regularizations throughout the entire network to learn features that are representative, robust, and more discriminative. Specifically, we introduce a pair of complementary attention modules, focusing on channel aggregation and position awareness, respectively. Then, we plug in a novel orthogonality constraint that efficiently enforces diversity on both hidden activations and weights. Through an extensive set of ablation study, we verify that the attentive and diverse terms each contributes to the performance boosts of ABD-Net. It consistently outperforms existing state-of-the-art methods on there popular person Re-ID benchmarks. * This also exploits attention mechanisms.• This is with a ResNet-152 backbone. This is with a DenseNet-121 backbone. ‡ Official codes are not released. We report the numbers in the original paper, which are better than our re-implementation.comes 3.40% for top-1 and 6.40% for mAP. We also considered SVDNet [13] and HA- CNN [50] which also proposed to generate diverse and uncorrelated feature embeddings. ABD-Net surpasses both with significant top-1 and mAP improvement. Overall, our observations endorse the superiority of ABD-Net by combing "attentive" and "diverse". VisualizationsAttention Pattern Visualization: We conduct a set of attention visualizations * * on the final output feature maps of the baseline (XE), baseline (XE) + PAM + CAM, and ABD-Net (XE), as shown in Fig.5. We notice that the feature maps from the baseline show little attentiveness. PAM + * * Grad-CAM visualization method [73]: https://github.com/ utkuozbulak/pytorch-cnn-visualizations; RAM visualization method [74] for testing images. More results can be found in the supplementary.

Discriminative Spatial Feature Learning for Person Re-Identification

“…Specifically, they aim to learn the features by improving loss functions [9,14,22,31,41,42,50,55,63], improving the training techniques [1,4,12,24,32,35,37,54], adding additional network modules [23,23,51,62], using extra semantic annotations [30,46,47,79] or generating more training samples [17,33,72,76,77]. Besides, more recent studies [3,6,8,10,27,28,38,46,48,53,58,61,64,65,67,80] integrate attention mechanisms into deep models to enhance the feature representation. To obtain the holistic features, most of these methods utilize global average pooling (GAP), global max pooling (GMP) or both of them on each channel of the feature map.…”

“…In the spatial feature branch, a SC-STN is firstly employed to refine the feature map, and then a VCN is introduced to extract the spatial feature. For both branches, the label-smoothed cross-entropy loss [8,32,53] and the ranked list loss [50] are utilized to make the features discriminative.…”

Person Re-Identification Using Additive Distance Constraint With Similar Labels Loss

“…The performance of our proposed person Re-ID method is compared with the state-of-the-art methods on both Market-1501 and DukeMTMC-reID datasets. The employed comparison methods include AlignedReID [45], IDE (ID-discriminative embedding) [39], SVDNet (singular vector decomposition net) [46], TriNet (triplet net) [30], Pyramid [47], AWTL (adaptive weighted triplet loss) [48], ABD-Net (Attentive but Diverse Net) [49], DSA-reID (Densely Semantically Aligned reID) [50], the baseline method in [51], and the baseline method together with triplet loss [51].…”

“…In practical industry applications that require simple but effective solutions, our proposed method could be one of the promising counterbalanced solutions to real-world person ReID tasks. [39] 79.5% 59.5% --SVDNet [46] 82.3% 62.1% 76.7% 56.8% TriNet [30] 84.9% 69.1% --Pyramid [47] 92.8% 82.1% --AWTL [48] 89.5% 75.7% 79.8% 63.4% ABD-Net [49] 95.6% 88.3% 89.0% 78.6% DSA-reID [50] 95.7% 87.6% 86.2% 74.3% Baseline [51] 93 When comparing the performance of our proposed ADCSLL and ADCSLL + triplet loss, the results show that ADCSLL together with triplet loss achieves better performance than ADCSLL alone. The rank-1 accuracies when using ADCSLL + triplet loss on Market-1501 and DukeMTMC-reID are 95.0% and 88.6%, respectively, 0.2% and 1.1% higher than the numbers when using ADCSLL alone.…”

Person Re-Identification by Discriminative Local Features of Overlapping Stripes

“…The deep learning model automatically decides through backpropagation what features to be extracted. Various deep neural models reported in the literature [6,32,33] can re-identify the individual in the presence of extreme distortion. Various pre-defined models that are AluxNet, Caffenet, Googlenet, VGG networks, ResNet, and SVDnet have also been used as feature extraction criteria for the person Re-ID.…”