Deep Residual Network and Transfer Learning-based Person Re-Identification
Arpita Gupta, Pratik Pawade, Ramadoss Balakrishnan
Abstract
Person re-identification (person re-id) is a field in which a person is recognized from different views, which plays a significant role in surveillance. One of the major problems in-person re-id is the unavailability of large labeled datasets, which has bounded the performance of the deep learning models. In this study, different variations were experimented with of deep residual networks trained with transfer learning, and then fine-tuning was done on the pre-trained model. These models are pre-trained on a larger dataset, ImageNet, to learn the visual features better. The proposed model is based on deep residual networks because of its better understanding and performance of the visual features, which leads to better classification. The proposed model was evaluated on the Market-1501 dataset, which consists of 1501 identities collected from 6 cameras. In this paper, the effect of hyperparameters is studied for better accuracy. The model has achieved the highest mAP score of 68.4%, with an improvement of 5.3%, and 96.0% of Rank-1 with 12.4%. The proposed model has outperformed all the existing models trained in a supervised manner on the Market 1501 dataset. The results have proved that this model could help better person re-id classification problems even when there is no availability of large labeled datasets and could be used for better security and surveillance.