Efficient Feature Extraction for Person Re-Identification via Distillation

Person re-identification has received increasing attention due to the high performance achieved by new methods based on deep learning. With larger networks of cameras being deployed, more surveillance videos need to be parsed, and extracting features for each frame remains a bottleneck. In addition, the feature extraction needs to be robust to images captured in a variety of scenarios. We propose using deep neural network distillation for training a feature extractor with a lower computational cost, while keeping track of its cross-domain ability. In the end, the proposed model is three times faster, without a decrease in accuracy. Results are validated on two popular person re-identification benchmark datasets and compared to a solution using ResNet.