Single-Shot Person Re-Identification based on Unsupervised Saliency Information

Resumen

Person re-identification task is important in video surveillance to improve security in public place. In recent years there is a lot of investigation about this thematic. However, the performance in these algorithms is affected by different problems in the scenes, for example, complex background, atmospheric conditions, etc. Some methods as deep learning and saliency descriptor have been used to solve these problems in the real world. In this paper, we developed a method based on the combination of convolutional neural network without fine-tuning and a saliency descriptor to weight all the information present into a person image.  Feature  maps  are  extracted  from  the  last  convolutional  layer of a neural  network  and  merged  with  other  salient  map  obtained  in spatial domain. Finally, different features are generated based on color histograms and local binary patterns. To verify the effectiveness of our proposal, the method is validated using VIPeR dataset and compared with others state of the art algorithms. The results shown that our proposal is easy to implement and is comparable with other approach using the Cumulative Matching Characteristic curve.