| Abstract [eng] |
In this research two different deep-learning person re-identification models are presented with their implementation, trained on three benchmark datasets: IUST_PersonReID, Market-1501, MARS. Market-1501 and MARS datasets were split according to the official data split for results to be comparable to existing research and an unofficial custom cross-camera split without junk to display best-case scenario how models would perform using clean data. First implementation trained on IUST_PersonReID dataset uses a GooGleNet Inception v1 backbone with weighted contrastive loss consisting of Gaussian-scored positive pair, margin-based negative pairs, identity-agnostic attention. Second implementation was trained on Market-1501 and MARS datasets separately producing two models using a DINOv2 feature extraction backbone with a combined loss consisting of AMSoftmax, circle loss, batch-hard triplet loss and a novel suggested temporal stripe consistency loss consisting of intra-frame and inter-frame components making sure horizontal stripe features are consistent across multiple frames across the tracklet. In this paper the evaluation results are reported as CMC curves, Rank-1 curve which show how Rank 1 is influenced by input/gallery sequence lengths, Rank-1/mAP tables without re-ranking to display out-of-the-box performance, with re-ranking to show optimal performance and lastly actual queries to showcase where model excels and struggles with predictions. |
