A Two-Stage Attention Based Efficient Second-Order Minimization Network for Planar Object Tracking

Planar object tracking, which detects the posture change of the 2D objects in a frame-by-frame manner, plays a critical role in various practical applications including visual simultaneous localization and mapping, visual servoing, and augmented reality. However, in real world, current methods suffer from ubiquitous challenges such as motion blur, texture degradation, illumination change, and partial occlusion. In this article, we propose a new tracking pipeline to handle the preceding challenges. The main novelty of our method lies in two new modules. Concretely, we first propose a Perspective Matrix Selection Module (PMSM) to rectify the search range error caused by large motion displacement. Second, we design an efficient attention mask module to highlight the target object while mitigating the disturbance caused by scene occlusion. The proposed method is evaluated on the challenging dataset POT. Experimental results show that the proposed method outperforms the state-of-the-art methods by a large margin in terms of accuracy and robustness over various tracking challenges, such as rotation, scale change, occlusion, motion blur, and perspective degradation.