The Segmentation Tracker With Mask-Guided Background Suppression Strategy

Segmentation-based tracking is currently a promising tracking paradigm with pixel-wise information. However, the lack of structural constraints makes it difficult to maintain excellent performance in the presence of background interference. Therefore, we propose a Segmentation tracker with mask-guided background suppression strategy. Firstly, a mask-aware module is designed to generate more accurate target masks. With the guidance of regression loss, features were selected that are sensitive only to the target region among shallow features that contain more spatial information. Structural information is introduced and background clutter in the backbone feature is suppressed, which enhances the reliability of the target segmentation. Secondly, a mask-guided template suppression module is constructed to improve feature representation. The generated mask with clear target contours can be used to filter the background noise, which increases the distinction between foreground and background of which. Therefore, the module highlights the target area and improves the interference resistance of the template. Finally, an adaptive spatiotemporal context constraint strategy is proposed to aid the target location. The strategy learns a region probability matrix by the object mask of the previous frame, which is used to constrain the contextual information in the search region of the current frame. Benefiting from this strategy, our method effectively suppresses similar distractors in the search region and achieves robust tracking. Broad experiments on five challenge benchmarks including VOT2016, VOT2018, VOT2019, OTB100, and TC128 indicate that the proposed tracker performs stably under complex tracking backgrounds.