Spectral Context Matching for Video Object Segmentation Under Occlusion

Abstract

Although numerous algorithms have been proposed for video object segmentation, it is still a challenging problem to segment video object in the case of occlusion. Video object localization is a critical step for an accurate object segmentation. To obtain an initial localization, we propose a new method, Spectral Context Matching (SCM), for a coarse object location. SCM rebuild the affinity Matrix using context information as similarity constraints of features to detect the corresponding areas. Adding with color and optical flow information, the initially estimated object location is selected. For object segmentation, we utilize a spatial-temporal graphical model on the estimated object region to get an accurate segmentation. In addition, we also impose an online update mechanism to detect and handle occlusion adaptively. Experimental results on DAVIS dataset and comparison with the-state-of-the-art method show that our proposed algorithm can efficiently handle heavy occlusion.