Part Context Learning for Visual Tracking

Context information is widely used in computer vision for tracking arbitrary objects[1, 3, 4]. Global context cannot deal with the object deformation problem, while the local part context interactions are relatively stable. When the target appearance changes gradually, the intrinsic property of internal interaction between the parts inside object and context interaction between object and background are relatively stable in spatiotemporal 3D space of tracking. To explore the structure property and stable relationship for overcoming complex environments, we propose a novel part context tracker. The Part Context Tracker (PCT) consists of an appearance model, an internal relation model and an context relation model. The internal relation model formulates the temporal relations of the object itself or the in-object parts themselves and the spatio-temporal relations between the object and inobject parts. The context relation model constructs the spatio-temporal relations between the in-object parts and the context parts and the temporal relations of the context parts themselves. Hence the physical properties and the appearance information are considered in the optimization process through parts and relations. The contributions are as follows: (1) We first propose a unified context framework which formulates the single object tracking as a part context learning problem. (2) The in-object parts and context parts are selected so that we not only pay attention to the appearance of object, but also focus on the relations among the object, the in-object parts and the context parts. (3) A simple yet robust update strategy using median filter is utilized, thereby enabling the tracker to deal with appearance change effectively and alleviate the drift problem. Our framework not only models the object with in-object parts, but also incorporates the interaction between the object and background with context parts. The deformable configuration [2, 5] together with the temporal structure of these parts are also considered in. In Fig. 1, with the object bounding box as the root R, the in-object parts I are defined as the parts selected inside R, which covers part of the object appearance. The context parts C are selected from the overlapping area between the object and the background. For a target with K in-object parts and M context parts, the configuration is denoted as B = (B0,B1...BK ,BK+1, ...,BK+M). Where B0 stands for the target bounding box R, (B1, ..,BK)∈ I are the K in-object part boxes, and (BK+1, ...,BK+M)∈ C are the M context part boxes. The corresponding features of the root and parts are represented as X = (x0, ...,xK ,xK+1, ...,xK+M). In a word, our framework models the object with three components: