Pose generation plays an essential role in computer graphics, such as game character design, and 3D modeling. Rather than inverse Kinematics solvers using deterministic heuristic methods suffering from poor diversity, sample-based methods promise to generate a wider variety of poses satisfying the given constraints. In order to obtain generative models from sample data, Generative Adversarial Networks (GANs) are widely used in many problems including pose generation. However, GANs are known to be suffering from mode collapse which causes the generation of specific patterns. Therefore, we propose a novel generative model for pose generation using Implicit Maximum Likelihood Estimation (IMLE), which is a training method for avoiding mode collapse by adaptive sampling of the input-output pairs. The proposed model accepts not only the latent variable, but also the condition of the pose such as a position of the kinematic model’s joint. We trained the proposed model by the IMLE’s optimization method using the dataset consisting of the pair of the pose condition and the corresponding joint angles. In the experiment of a simulated 3-DoF arm simulation, the proposed model successfully avoided mode collapse, thus better diversity rather than the GAN variants while satisfying the given conditional input. Furthermore, we report that the proposed model performs lower prediction error and higher variance than the GAN variants through the experiments on 30-DoF human pose using CMU Mocap Dataset.
抄録全体を表示