Adaptive Sample Selection Strategy for Training Object Trackers

Abstract

Most visual object tracker (VOT) models are trained in a supervised manner, and their performance is completely dependent on the quality of data labels. However, the sample selection criteria of existing region proposal network based trackers use heuristic rules and do not consider object characteristics of individual images. This hinders learning semantic features at the data level. In this thesis, we propose a training sample selection strategy using semantic mask for visual object tracker. The proposed algorithm selects balanced training samples by considering the shape of the object and the distribution of the candidate samples. Our method achieved 0.475 EAO on the VOT evaluation dataset and a success rate of 0.510 on the LaSOT benchmark, which exceeded the performance of the baseline model SiamRPN++.