A Multi-scale Triplet Deep Convolutional Neural Network for Person Re-identification

Person re-identification, aiming to identify images of the same person from non-overlapping camera views in different places, has attracted a lot of interests in intelligent video surveillance. As one of the newly emerging applications, deep learning has been incorporated into the feature representation of person re-identification. However, the existing deep feature learning methods are difficult to generate the robust and discriminative features since they use a fixed scale training and thus fail to adapt to diversitified scales for the same persons under realistic conditions. In this paper, a multi-scale triplet deep convolutional neural network (MST-CNN) is proposed to produce multi-scale features for person re-identification. The proposed MST-CNN consists of three sub-CNNs with respect to full scale, top scale (top part of persons) and half scale of the person images, respectively. In addition, these complementary scale-specific features are then passed to the l2-normalization layer for feature selection to obtain a more robust person descriptor. Experimental results on two public person re-identification datasets, i.e., CUHK-01 and PRID450s, demonstrate that our proposed MVT-CNN method outperforms most of the existing feature learning algorithms by 8%-10% at rank@1 in term of the cumulative matching curve (CMC) criterion.