End-to-end training of CNN ensembles for person re-identification
作者：
Highlights：
• An end-to-end ensemble learning method is proposed to address the overfitting problem in discriminative person ReID models.
• The presented approach is very efficient in both training and test times compared to the conventional ensemble methods.
• We avoid custom design of network architecture specialized to ReID and make minimal changes in DenseNet architecture.
• We obtain accurate and diverse base learners; their combined features improves Rank-1 and mAP scores significantly.
• We achieve state-of-the-art results on three large scale benchmark datasets.
摘要
•An end-to-end ensemble learning method is proposed to address the overfitting problem in discriminative person ReID models.•The presented approach is very efficient in both training and test times compared to the conventional ensemble methods.•We avoid custom design of network architecture specialized to ReID and make minimal changes in DenseNet architecture.•We obtain accurate and diverse base learners; their combined features improves Rank-1 and mAP scores significantly.•We achieve state-of-the-art results on three large scale benchmark datasets.
论文关键词：Deep networks,Ensemble learning,Person re-identification
论文评审过程：Received 21 March 2019, Revised 26 December 2019, Accepted 28 February 2020, Available online 4 March 2020, Version of Record 18 March 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107319
