Segmentation tracker based on SiamMask for inspection video

With the development of the computer vision, intelligent inspection of power equipment has become an important task for power system. In this paper, we propose a segmentation tracker to perform video object tracking and video object segmentation for intelligent inspection at the same time in real time. We design a segmentation verification function to judge the mask from three perspectives including the mask quality, mask-to-box ability and distance penalty to improve the mask accuracy for the power equipments. We also design a module to use the inter-frame information to propagate the mask as a supplement for the SiamMask in which only the first frame information is used. The experiments both on the benchmark dataset and the inspection videos show great performance of our algorithm.