Person re-identification via integrating patch-based metric learning and local salience learning
作者:
Highlights:
• We propose an extendable re-id framework, which contains two related parts: patch-based metric learning and local salience learning. First, to handle the problem of pose variant, CNN features are extracted to represent the person, and then a light patch-based metric learning method-pLMNN is proposed to enhance the discriminative ability of raw features.
• We propose a Kmeans-based local salience learning algorithm to train the weights of image patches. Meanwhile, a general similarity computation scheme is presented to relieve an existing training problem, i.e., the parameters need to be re-trained for different datasets.
• The experimental results on two challenging datasets demonstrates the effectiveness of our method.
摘要
•We propose an extendable re-id framework, which contains two related parts: patch-based metric learning and local salience learning. First, to handle the problem of pose variant, CNN features are extracted to represent the person, and then a light patch-based metric learning method-pLMNN is proposed to enhance the discriminative ability of raw features.•We propose a Kmeans-based local salience learning algorithm to train the weights of image patches. Meanwhile, a general similarity computation scheme is presented to relieve an existing training problem, i.e., the parameters need to be re-trained for different datasets.•The experimental results on two challenging datasets demonstrates the effectiveness of our method.
论文关键词:Person re-identification,CNN feature,Patch-based metric learning,Local salience learning,Cross-dataset
论文评审过程:Received 15 November 2016, Revised 12 March 2017, Accepted 19 March 2017, Available online 24 March 2017, Version of Record 21 November 2017.
论文官网地址:https://doi.org/10.1016/j.patcog.2017.03.023