Homography Decomposition Networks for Planar Object Tracking

Planar object tracking plays an important role in AI applications, such as robotics, visual servoing, and SLAM. Although the previous planar trackers work well most scenarios, it is still a challenging task due to rapid motion large transformation between two consecutive frames. The essential reason behind this problem that condition number of non-linear system changes unstably when searching range homography parameter space becomes larger. To end, we propose novel Homography Decomposition Networks~(HDN) approach drastically reduces stabilizes by decomposing into groups. Specifically, similarity estimator designed predict first group robustly deep convolution equivariant network. By taking advantage scale rotation estimation with high confidence, residual estimated simple regression model. Furthermore, proposed end-to-end network trained semi-supervised fashion. Extensive experiments show our outperforms state-of-the-art methods at margin on POT, UCSB POIC datasets. Codes models are available https://github.com/zhanxinrui/HDN.