| Abstract: |
We propose a novel multi-scale Siamese architecture to perform person re-identification using deep learning. The scenario considered in this work is similar to that found in movie/concert halls, where persons enter in a queue one-by-one through the entry gates and leave in a similar way through the exit gates. Effectiveness of Siamese network based re-identification is evident from the recent research work in this domain. Here, we focus on improving the accuracy of the existing re-identification techniques by introducing different dilation rates in the convolution layers of the Siamese network, thereby enabling capturing of detailed visual features. We also introduce a silhouette part-based analysis to preserve the spatial relationships among the different silhouette segments at a high resolution. The proposed Siamese network model has been fine-tuned through cross-validation and the pre-trained network has been made available for further comparison. Rigorous evaluation of our approach against varying training parameters, as well as comparison with state-of-the-art methods over four popularly used data sets, namely, CUHK 01, CUHK 03, Market1501, and VIPeR, verify its effectiveness. |