Abstract: With the emergence of 3D virtual worlds, 3D social media, and massive online games, the need for diverse, high-quality, animation-ready characters and avatars is greater than ever. To animate characters, artists hand-craft articulation structures, such as animation skeletons and part deformers, which require significant amount of manual and laborious interaction with 2D/3D modeling interfaces. This thesis presents deep learning methods that are able to significantly automate the process of character rigging.
First, the thesis introduces RigNet, a method capable of predicting an animation skeleton for an input static 3D shape in the form of a polygon mesh. The predicted skeletons match the animator expectations in joint placement and topology. RigNet also estimates surface skin weights which determine how the mesh is animated given the different skeletal poses. In contrast to prior work that fits pre-defined skeletal templates with hand-tuned objectives, RigNet is able to automatically rig diverse characters, such as humanoids, bipeds, quadrupeds, toys, birds, with varying articulation structure and geometry. RigNet is based on a deep neural architecture that directly operates on the mesh representation. The architecture is trained on a diverse dataset of rigged models that we mined online and curated. The dataset includes 2.7K polygon meshes, along with their associated skeletons and corresponding skin weights.
Second, the thesis introduces Morig, a method that automatically rigs character meshes driven by single-view point cloud streams capturing the motion of performing characters. Compared to RigNet, this method is not only motion-aware, but also able to animate the 3D meshes according to the captured point cloud motion. MoRig's neural network encodes motion cues from the point clouds into features that are informative about the articulated parts of the performing character. These motion-aware features guide the inference of an appropriate skeletal rig for the input mesh, which is then animated based on the point cloud motion. Morig can also rig and animate diverse characters, and accounts for occluded regions in the point clouds and mismatches in the part proportions between the input mesh and captured character.
Third, the thesis introduces APES, a method that takes as input 2D raster images depicting a small set of poses of a character shown in a sprite sheet, and identifies articulated parts useful for rigging the character. APES uses a combination of neural network inference and integer linear programming to identify a compact set of articulated body parts, e.g. head, torso and limbs, that best reconstruct the input poses. Compared to RigNet that requires a large collection of training models with associated skeletons and skinning weights, APES' neural architecture relies on supervision from (i) pixel correspondences readily available in existing large cartoon image datasets (e.g., Creative Flow), (ii) a relatively small dataset of 57 cartoon characters segmented into moving parts.
Finally in the thesis, I discuss my future work that will use rig to control implicit field, such as NeRF, to render articulated human avatars. By involving differentiable volumetric rendering, the method will gain supervision from 2D observations, and produce photo-realistically rendered animation controlled by user-specified poses and views.
Advisor: Evangelos Kalogerakis