Investigating Domain Transfer and Viewpoint in the Context of Person Re-Id

Abstract

Deep learning has significantly improved Re-Id per- formance but it requires a large amount of data, however, obtaining data is expensive from both time and money perspective. Inspired by ImageNet pre- trained models and synthetic data generation techniques, this paper investigates to utilise real-world and syn- thetic Re-Id datasets to augment task performance. Firstly, we propose two methods to apply external Re-Id data, NDTL (Neighbour-Domain Transfer Learning) and NDDS (Neighbour-Domain Data Stitching). Secondly, we quantitatively illustrate that both real-world and syn- thetic data could mitigate Re-Id data shortage problems, using Re-Id dataset to pre-train models is better than us- ing ImageNet, we achieve up to 28.2% mAP improvement on DukeMTMC and 5.2% on Market-1501. Moreover, we find out that viewpoint, one of Re-Id relevant factors, has the an influence on the system performance due to viewpoint-wise non-alignment and unbalance of the orig- inal dataset, it also assists the performance if train set is augmented balanced. Our research strongly illustrates both real-world and synthetic Re-Id dataset can effec- tively augment Re-Id task, viewpoint is an essential fac- tor and based on which, train-test distribution dramati- cally influences Re-Id performance, and balancing train classes are also helpful to improve the performance.