Enhance tracker's classifier via view morphing
Xiang Li, Yue Fei Zhou · 2017
Under the situation of model-free tracking, classifier in tracking system often suffers from limited input samples and bounded observation points. System would perform better, provided as much information of the target as possible. However, it would be extremely hard to select a perfect feature extractor in all cases. In this paper, a novel way to dig out more information from existent pictures is introduced, without changing the feature extractor. A brand new module, called View Morphing, was installed on tracking system. It can use several basic target images to generate lots of virtual samples, even the camera parameters are unavailable. These morphed pictures described the target at various observation points, contained naive 3D constrains, and would help the classifier understand the target better. Experiments showed that view morphing module enhanced both classical tracker and deep-learning one. The brand new module provided a possible solution on lack of training images.