Deep Feature Learning with Relative Distance Comparison for Person Re-identification

• DOI: 10.48550/arxiv.1512.03622
• Publication Date: 2015-01-01
• Authors: Shengyong Ding; Liang Lin; Guangrun Wang; Hongyang Chao
• Keywords: Feature (linguistics); Similarity (geometry); Identification

Identifying the same individual across different scenes is an important yet difficult task in intelligent video surveillance. Its main difficulty lies how to preserve similarity of person against large appearance and structure variation while discriminating individuals. In this paper, we present a scalable distance driven feature learning framework based on deep neural network for re-identification, demonstrate its effectiveness handle existing challenges. Specifically, given training images with class labels (person IDs), first produce number triplet units, each which contains three images, i.e. one matched reference mismatched reference. Treating units as input, build convolutional generate layered representations, follow $L2$ metric. By means parameter optimization, our tends maximize relative between pair unit. Moreover, nontrivial issue arising that organization cubically enlarges triplets, image can be involved into several units. To overcome problem, develop effective generation scheme optimized gradient descent algorithm, making computational load mainly depends original instead triplets. On challenging databases, approach achieves very promising results outperforms other state-of-the-art approaches.