AFOD: Adaptive Focused Discriminative Segmentation Tracker

Visual object tracking is a fundamental task in computer vision which could be integrated into numerous real-world applications. Traditional object tracking methods focus on providing the bounding box as object position, while some recent trackers start to consider the combination of segmentation module to generate the binary segmentation mask, pursuing more accurate localization. However, how to effectively integrate different information for accurate and robust tracking is an open question. In this paper, we propose a novel Adaptive FOcused Discriminative (AFOD) segmentation tracker with the following advanced components. For localization, a more discriminative light weight online target appearance model is employed to provide robust position estimation. For segmentation, leveraging the backbone semantic feature, the coarse segmentation feature, and the localization feature, an offline trained fine segmentation model with IoU optimization is utilized to generate the accurate high resolution masks. The boundary detection further enhances the segmentation quality. For combination, an adaptive prediction strategy is proposed to better integrate the information from two types of predictions, i.e. box and mask. AFOD achieves leading performance on two tracking benchmarks including the bounding box annotated VOT2018 and the segmentation mask annotated VOT2020, while running close to real-time.