Correspondence construction for cartoon animation via sparse coding

In 2D animation, it is a tedious task and time-consuming that drawing in-betweens for the animator. Correspondence construction between two key frames is a necessary condition for auto-inbetween in the computer animation auxiliary system. In this paper, we combine patch alignment framework (PAF) with the idea of sparse coding for correspondence construction. Specifically, local patches construction can have a large impact on the accuracy of correspondence. Therefore, in our framework, in order to construct local patches in each point on an object and align these patches in a new feature space, we adopt sparse coding instead of k-nearest neighbor method in patch construction. The correspondences between two objects can be detected by subsequent clustering method. This approach can efficiently improves the performance of correspondence construction. To optimize the proposed framework, we use least angle regression (LARS) method to overcome the slow operating efficiency problem of lasso. Experimental results on our cartoon data set which is built on industrial production suggest the advanced accuracy of correspondence construction in our improved framework, and is even better than the framework of using k-nearest neighbor algorithm to construct local patches.