标签：
Discriminative Learning Neural Network for Person Re-identification
Hao Sheng, Meiyuan Liu, Yan Huang, Yanwei Zheng, Xiong Zhang
Abstract:
Person re-identification across disjoint camera views is an important yet tough task in intelligent video surveillance due to the large variations of lightings, poses, image resolutions and cluttered backgrounds across camera views. In this paper, we present a discriminative learning neural network (DLNN) based on convolutional neural network to address the problem of pedestrian re-identification. Given a pair of images as input, our network learn the similarity directly from the raw image pixels by jointly handling the misalignment, photometric transforms and background clutter, as a result of which produce a score indicating whether the two images depict the same person.
Unlike the existing neural networks applied to person re-identification which extract features from each input image respectively before estimate similarity on these features, we innovatively concatenate the 3 color channels of the two pictures to a veritably input of 6 channels, and then construct a hierarchical feature representation by three subsequent convolution layers, which is proved to be of vital importance in the face of re-identification by our experiment. On the most challenging dataset CUHK03, we achieve a comparable result to the state-of-the-art and outperform all the other approaches, while our model is of half size of the model used by the state-of-the-art.
标签：
原文地址：http://www.cnblogs.com/sylviaDollar/p/5341184.html
