Optimization approaches applied to animation setup transfer for visual effects

摘要

In today’s animation and visual effects (VFX) industry, time is precious. Making a 3D character ready for animation, and by ready we mean preparing the skeletons and skin weights, is a very time-consuming task. Thus speeding up this iterative creation process will be useful. This thesis is part of a broader project done in the Multimedia Laboratory at École de technologie supérieure (ÉTS) in collaboration with an industrial partner, Autodesk, and also with contributions of a postdoctoral fellow, two PhD students, and a professional’s master student. The main goal of this project is to transfer the already-done artist’s work to other 3D characters. This process is done through a pipeline consisting of a geometric correspondence computation, skeleton transfer, and weight transfer. Within the broader project, this dissertation makes specific contributions to the analysis of underlying mathematical models and to enhance and optimize the skeleton retargeting.ududTo transfer the animation setup, first a geometric correspondence based on a set of markers is computed between the source mesh and target mesh. It should be mentioned here that the whole process of skeleton retargeting and skin weights transferring does not rely on a specific geometric correspondence method. Each joint can be extracted as a blended combination of its weighted vertices. The weighted vertices of each joint are a set of vertices, which are affected by the transformation of the joint. The computed vertex-to-point correspondence is used to position the source joint within the target mesh, in a way that reproduces the spatial relationship found between the source joint and its weighted vertices. In this dissertation, two different approaches are developed for locating the joints within the target mesh. One approach is done by solving an energy minimization problem, and the other one is done using generalized Procrustes analysis. The skeleton retargeting results achieved using these two approaches are proposed and compared, mostly in the matter of speed and simplicity. Furthermore, the rotation and orientation of each joint need to be retargeted as well. The orientation is copied from the source joints’ orientation directly. Transferring the rotation is done by finding the rotation alignment between the point cloud of the joint’s weighted vertices on the source and their corresponding locations on the target. A pose normalization procedure is designed to make sure that the joints in the source and target share the same limb orientation. To improve the skeleton quality, the options for aligning the spine and mirroring the joint positions to take advantage of symmetrical mesh’s features, are also developed. At last, the skin weights are transferred from the source mesh to the target mesh.ududThe approaches are tested on a diversity of source and target characters with a vast range of mesh topology and morphology in both inverse and forward kinematics animations. This method allows for reusing existing animation setup and reducing the manual interventions.