Efficiency-enhanced Progressive Sampling Method on One-shot Person Re-Identification

Label estimation on unlabeled samples has become a typical method in one-shot Person Re-Identification. Many existing methods take this form of data augmentation to create the conditions for fully supervised learning. The state-of-theart method uses a cross-iterative mode and achieves huge performance improvement. However, the learning efficiency is very low in this mode. In this paper, we first propose an exponential sampling curve to further explore the relations between sampling number and the performance of the model in each iteration. We found that expanding a proper amount of pseudo-label samples for training can accelerate the growth of performance at the risk of losing final performance. To avoid performance loss, an efficiency-enhanced progressive sampling method is proposed subsequently to expanding sampling number in each iteration by amplification factor and increment factor. Our method not only makes full use of more pseudo-label samples, but also avoids adding too many mislabeled samples at the beginning. Our method is validated on DukeMTMCVideoReID dataset, with the results that our method has a performance comparable to the state-of-the-art method and reduces the number of training iterations by 30% to obtain the best performance.