Towards Robust Direction Invariance in Character Animation

Abstract

In character animation, direction invariance is a desirable property. That is, a pose facing north and the same pose facing south are considered the same; a character that can walk to the north is expected to be able to walk to the south in a similar style. To achieve such direction invariance, the current practice is to remove the facing direction's rotation around the vertical axis before further processing. Such a scheme, however, is not robust for rotational behaviors in the sagittal plane. In search of a smooth scheme to achieve direction invariance, we prove that in general a singularity free scheme does not exist. We further connect the problem with the hairy ball theorem, which is better‐known to the graphics community. Due to the nonexistence of a singularity free scheme, a general solution does not exist and we propose a remedy by using a properly‐chosen motion direction that can avoid singularities for specific motions at hand. We perform comparative studies using two deep‐learning based methods, one builds kinematic motion representations and the other learns physics‐based controls. The results show that with our robust direction invariant features, both methods can achieve better results in terms of learning speed and/or final quality. We hope this paper can not only boost performance for character animation methods, but also help related communities currently not fully aware of the direction invariance problem to achieve more robust results.

Discovering diverse athletic jumping strategies

“…In principle, we could also use direction-invariant features as in DeepMimic, and include the relative transformation to the obstacle into the feature set. However, as proved in [Ma et al 2019], there are no direction-invariant features that are always singularity free. Direction-invariant features change wildly whenever the character's facing direction approaches the chosen motion direction, which is usually the global up-direction or the 𝑌 -axis.…”

A Hierarchy-Aware Pose Representation for Deep Character Animation

“…Even with abundant weights and extensive training, they still occasionally produce big errors. In previous work, the efficiency of deep learning approaches has been found to be highly dependent upon the quality of the dataset and the representation used [Ma et al 2019;Xiang and Li 2020;Zhou et al 2019], due to the fact that the network learns motion patterns from the data. A rich encoding of motion, based on a concrete mathematical model, encourages the network to properly disentangle complex characteristics of motion which are present in the data.…”

Improving Skeleton-based Action Recognitionwith Robust Spatial and Temporal Features

“…Direction Invariant Features Within the skeleton-based action recognition literature, some works use raw joint positions as input [42]; some transform joint positions to a semi-local coordinate system which is close to character [29,31,14]. We adopt the principled way of transforming input features into Direction Invariant Features(DIF) described in [24]. Our experiments show that this simple preprocessing alone can boost performance by a large margin.…”

“…Various input features have been employed in the literature, such as global joint positions [42], and joint positions transformed into a semi-local coordinate system where the origin is defined by the character's position in the first frame of the clip, and the axes defined by the shoulder joints and the gravitational direction [31]. We follow the principled way of calculating Direction Invariant Features (DIF) proposed in [24], which outperforms global features for various tasks in character animation.…”

“…That is, a walk to the south and a similar walk to the north should be classified as the same type of motion. It is well-known in the computer animation community that direction invariance is desirable [24]. However, most action recognition systems in the computer vision community use absolute joint coordinates as input, so motions towards different directions look quite different to the neural networks.…”