Deformable and Occluded Object Tracking via Graph Learning

Object deformation and occlusion are ubiquitous problems for visual tracking. Though many efforts have been made to handle object deformation and occlusion, most existing tracking algorithms fail in case of large deformation and severe occlusion. In this paper, we propose a graph learning-based tracking framework to handle both challenges. For each consecutive frame pair, we construct a weighted graph, in which the nodes are the local parts of both frames. Our algorithm optimizes the graph similarity matrix until two disconnected subgraphs separate the foreground and background nodes. We assign foreground/background labels to the current frame nodes based on the learned graph and estimate the object bounding box under an optimization framework with the predicted foreground parts. Experimental results on the Deform-SOT dataset shows that the proposed method achieves the state-of-the-art performance.