Light-Weight Visual Feature Based Labeling (LVFL) for Unsupervised Person Re-identification

Abstract

Person re-identification is a vital problem in smart video surveillance environment. Performing person reidentification with an unlabeled dataset is challenging. Even though certain labeling mechanisms are available in the literature, the computation overhead prevents the system to perform re-identification task dynamically. To overcome this issue, we propose a Light-weight Visual Feature based Labeling (LVFL) method to label the person re-identification images and reduce the computation overhead than the state-of-themethods. The computation overhead is reduced at three stages namely model initialization, neural network utilization and algorithmic complexity through evaluation of the cluster quality and Cumulative Match Curve (CMC) scores. The proposed method reports a reduced computation complexity than the traditional unsupervised person re-identification methods by determining a tight bound fine-tuning with a very less CMC score trade-off. Experimental results tested on three major benchmark datasets namely DukeMTMC re-id, Market1501 and CUHK03 show that the proposed LVFL produces a decent matching performance with a computation overhead reduction of about 29 % to 41 %.