PaaRPN: Probabilistic anchor assignment with region proposal network for visual tracking

Abstract

Recently, visual trackers based on region proposal networks (RPN) have attracted widespread attention due to their relatively high efficiency and excellent performance. RPN-based trackers mainly combine a classification branch and a regression branch to predict a target’s state. These branches are all under the guidance of pre-defined anchor boxes. RPN-based trackers, however, first compute the Intersection-over-Union (IoU) between the anchor boxes and ground truth boxes, and then use a fixed IoU threshold to separate negative and positive training samples. The limit of this design lies in the fact that these trackers lack an analysis of the actual content of the intersecting regions, which may include distractor objects or few meaningful regions of the tracked target. In this research, we propose a probabilistic anchor assignment with region proposal network (PaaRPN) that can adaptively separate anchors into negative samples and positive samples according to the model’s current learning status. To this end, we first calculate the classification scores of the anchor boxes conditioned on the current model and fit a probability distribution to the classification scores. The whole tracking model is then trained with anchor boxes separated into negative and positive samples in a probabilistic manner. Moreover, we introduce an online learning method in the PaaRPN framework that enables the model to have powerful discriminative abilities by exploiting both background and target appearance information. We tested the PaaRPN tracker on six tracking benchmarks to exhibit the effectiveness of the proposed method. In particular, our model outperforms a strong RPN tracker, SiamRPN++, with AUC scores improvements of 0.613 → 0.657 and 0.496 → 0.565 on UAV123 and LaSOT, respectively.