Part-based Feature Extraction for Person Re-identification
Cheng‐Lin Liu, Tianlong Bao, Ming Bo Zhu · 2018
In this paper, we propose a new part-based CNN feature extraction method for end-to-end person re-identification. In our method, the input images are first divided into two different non-overlapping parts, and then two different CNN models are trained for classifying them respectively. Next, in testing phase, we use these two trained CNN models to extract features from different parts of persons separately and then combine them together to construct the overall features of the input images. Finally, metric is utilized to rank the distances between the overall features of the different persons. To make maximize performance, we also try different types of CNN models with various distance metrics. Experiment on dataset Market-1501 demonstrates that our proposed method for person re-identification significantly outperforms state-of-the-art models. And experiment on dataset MARS shows that our method can also get competitive performance on large dataset.