Search for a command to run...
In this work we propose a new architecture for person re-identification. As\nthe task of re-identification is inherently associated with embedding learning\nand non-rigid appearance description, our architecture is based on the deep\nbilinear convolutional network (Bilinear-CNN) that has been proposed recently\nfor fine-grained classification of highly non-rigid objects. While the last\nstages of the original Bilinear-CNN architecture completely removes the\ngeometric information from consideration by performing orderless pooling, we\nobserve that a better embedding can be learned by performing bilinear pooling\nin a more local way, where each pooling is confined to a predefined region. Our\narchitecture thus represents a compromise between traditional convolutional\nnetworks and bilinear CNNs and strikes a balance between rigid matching and\ncompletely ignoring spatial information.\n We perform the experimental validation of the new architecture on the three\npopular benchmark datasets (Market-1501, CUHK01, CUHK03), comparing it to\nbaselines that include Bilinear-CNN as well as prior art. The new architecture\noutperforms the baseline on all three datasets, while performing better than\nstate-of-the-art on two out of three. The code and the pretrained models of the\napproach can be found at https://github.com/madkn/MultiregionBilinearCNN-ReId.\n