Construction of Diverse DropBlock Branches for Person Reidentification

Abstract

In this paper, we propose to use the data augmentation of batch drop-block with varying dropping ratios for constructing diversity-achieving branches in person re-identification. Since a considerable portion of input images may be dropped, this reinforces feature learning of the un-dropped region but makes the training process hard to converge. Hence, we propose a novel double-batch-split co-training approach for remedying this problem. In particular, we show that the feature diversity can be well achieved with the use of multiple dropping branches by setting individual dropping ratio for each branch. Empirical evidence demonstrates that the proposed method performs competitively on popular person Re-ID datasets, including Market-1501, DukeMTMC-reID and CUHK03, and the use of more dropping branches can further boost the performance. Source code is available at url.