Strong Baseline with Auto-encoder for Scale-Invariant Person Re-identification

Person Re-identification associates the same person images captured by disjoint cameras. The different training strategies and precise hyper-parameter selection are required for the construction of an effective person re-identification. However, in the current work, the tests are carried out by trial and error, which takes a long time to reach an optimal solution. We experiment with various training strategies with the metric learning model serving as the baseline to address this issue. Existing person re-identification benchmarks have insufficient samples for training the ReID model. Also, the previous approaches fail when the same individual appears in varied sizes. To solve this issue, an auto-encoder module is used to create a single image in three distinct scales to enhance the sample size and tackle the scale variation problem of the person reidentification. In addition, a performance comparison is made between the baseline and the baseline with auto-encoder to demonstrate the influence of an auto-encoder as a data augmentation for the person re-identification. The adoption of auto-encoder module and bestperforming training techniques to a baseline have enhanced the rank1 and mAP on Market1501, DukeMTMC-reID datasets.