Online Deformable Object Tracking Based on Structure-Aware Hyper-graph

Dawei Du, Honggang Qi, Wenbo Li, Longyin Wen, Qingming Huang, Siwei Lyu


   Recent advances in online visual tracking focus on designing part-based model to handle the deformation and occlusion challenges. However, previous methods usually consider only the pairwise structural dependencies of target parts in two consecutive frames rather than the higher-order constraints in multiple frames, making them less effective in handling large deformation and occlusion challenges. This paper describes a new and efficient method for online deformable object tracking. Different from most existing methods, our work exploits higherorder structural dependencies of different parts of the tracking target in multiple consecutive frames. We construct a structureaware hyper-graph to capture such higher-order dependencies, and solve the tracking problem by searching dense subgraphs on it. Furthermore, we also describe a new evaluating dataset for online deformable object tracking (the Deform-SOT dataset), which includes 50 challenging sequences with full annotations that represent realistic tracking challenges such as large deformations and severe occlusions. The experimental result of the proposed method shows considerable improvement in performance over the state-of-the-art tracking methods.


Illustrative Results


     • Deform-SOT Dataset [Project Page]          
     • SAT Matlab source code [Code].


     If you use the dataset, our tracking results or the source code, please cite our paper:
           Dawei Du, Honggang Qi, Wenbo Li, Longyin Wen, Qingming Huang, Siwei Lyu" Online Deformable Object Tracking Based on Structure-Aware Hyper-graph", IEEE Transaction on Image Processing (TIP), 2016. [PDF