Regressive Scale Estimation for Visual Tracking
Lutao Chu, Huiyun Li · 2019
In visual tracking, the aspect ratio variation of a target frequently appears and accurate scale estimation is challenging. Most tracking methods employ an exhaustive scale search to estimate the target sizes. Due to expensive computation, these methods can only search for finite and discrete scales. They fail if scale varies out of range. We propose a Regressive Scale Estimation (RSE) tracking framework by integrating a positioning model and bounding box regression, for position and scale estimation respectively. We generalize the original single-channel bounding box regression to multi-channel situations, to allow for better employment of multi-channel features. Contrary to most approaches, our method directly searches the target sizes within continuous scale space, which can predict any scale, not limited by a manually specified number of scales. Our scale model regresses the height and width of target simultaneously rather than treating them as one proportional scaling factor, leading to better flexibility. Experiments on Visual-Object-Tracking 2017 (VOT) benchmarks prove that our approach can adapt the target scale and achieve state-of-the-art performance in terms of accuracy and robustness.