Utilizing attention mechanism with exemplar memory for improving domain adaptive person re-identification

In an Unsupervised Domain Adaptation (UDA) task, extracted features from the entire image lead to a negative transfer of irrelevant knowledge. An attention mechanism may highlight the suitable transferable region of an image. Here in this paper, the representative power of the deep re-identification (re-ID) network of an exemplar memory-based domain adaptive unsupervised Person Re-identification (PRId) model is increased with our chosen optimal Bottleneck Attention Module (BAM). Performance enhancement of the deep re-ID network itself has been attempted by very few research works, suffering from a misinterpretation of global features or non-determination of local features due to the presence of occlusions, background interference, view-point variations, low camera resolution, etc. The proposed approach shows remarkable performance improvement in these specific challenging scenarios. Our attention-enabled unsupervised PRId model is evaluated in a cross-platform environment by labeled training in the DukeMTMC-reID dataset and unlabeled target testing in the Market-1501 dataset. The proposed method also enhances the overall performance of our baseline architecture. It provides a 2.12% relative increment in mean Average Precision (mAP) value, 0.93% increment in Top-1 value, 0.33% increment in Top-10 value, and 0.53% increment in Top-20 value respectively.