Conditioners for Adaptive Regression Tracking.

Ding Ma,Xiangqian Wu

DOI: https://doi.org/10.1007/978-3-030-88004-0_18

2021-01-01

Abstract:Given the input of the whole search area, the goal of regression tracking is to estimate the location of the target object by a calculated response map which usually generates by a Gaussian function. Due to object appearance variation caused by factors such as lighting change and deformation, the update is essential during tracking process. To this end, recent deep regression trackers usually fine-tune the model using hundreds of iterations of gradient descent. Although such trackers achieve pretty good performance, the fine-tuning process is inefficient and limits the tracker's speed which is a relatively important evaluation metric. To solve this problem, we propose an adaptive regression model that utilizes a single forward pass to adapt the model to the specific target object. During forward pass process, two additional networks are designed to adjust the parameters of intermediate layer of the regression model. Such two networks are referred to as conditioners which pay attention to the context and spatial information simultaneously. Extensive experiments illustrate that our approach is much faster than the fine-tuned trackers while achieving competitive performance on four public tracking benchmarks including OTB2015, TC128, UAV123 and VOT2018.