Deep Features for Person Re-identification

Matching observations captured by pedestrian detectors across the cameras with non-overlapping views, known as person re-identification, is challenging due to the appearance changes caused by pose, viewpoint and illumination variations, occlusions and cluttered background. Different from various hand-crafted features, this paper extract the features through the fine-tuned deep convolutional neural network to measure the similarities between the observations. Our approach significantly outperforms the state-of-the-art methods on the publicly available CUHK03 dataset.