Dominik Borer

Dominik Borer, Jakob Buhmann, Martin Guay

Modern pose estimation models are trained on large, manually-labelled datasets which are costly and may not cover the full extent of human poses and appearances in the real world. With advances in neural rendering, analysis-by-synthesis and the ability to not only predict, but also render the pose, is becoming an appealing framework, which could alleviate the need for large scale manual labelling efforts. While recent work have shown the feasibility of this approach, the predictions admit many flips due to a simplistic intermediate skeleton representation, resulting in low precision and inhibiting the acquisition of any downstream knowledge such as three-dimensional positioning. We solve this problem with a more expressive intermediate skeleton representation capable of capturing the semantics of the pose (left and right), which significantly reduces flips. To successfully train this new representation, we extend the analysis-by-synthesis framework with a training protocol based on synthetic data. We show that our representation results in less flips and more accurate predictions. Our approach outperforms previous models trained with analysis-by-synthesis on standard benchmarks.

Dominik Borer, Dominik Lutz, Martin Guay

One of the main challenges with embodying a conversational agent is annotating how and when motions can be played and composed together in real-time, without any visual artifact. The inherent problem is to do so---for a large amount of motions---without introducing mistakes in the annotation. To our knowledge, there is no automatic method that can process animations and automatically label actions and compatibility between them. In practice, a state machine, where clips are the actions, is created manually by setting connections between the states with the timing parameters for these connections. Authoring this state machine for a large amount of motions leads to a visual overflow, and increases the amount of possible mistakes. In consequence, conversational agent embodiments are left with little variations and quickly become repetitive. In this paper, we address this problem with a compact taxonomy of chit chat behaviors, that we can utilize to simplify and partially automate the graph authoring process. We measured the time required to label actions of an embodiment using our simple interface, compared to the standard state machine interface in Unreal Engine, and found that our approach is 7 times faster. We believe that our labeling approach could be a path to automated labeling: once a sub-set of motions are labeled (using our interface), we could learn a prediction that could attribute a label to new clips---allowing to really scale up virtual agent embodiments.