Dataset distillation seeks to compress a large training set into a small number of synthetic examples that still allow a model to learn effectively. For skeleton-based action recognition, the challenge is to preserve the motion variations that distinguish different actions while keeping the set small.
ProMoDD, described in a new arXiv preprint, applies this idea to 3D human motion using prototype-conditioned diffusion models. The method is designed to capture discriminative motion variation, which is essential for recognizing actions from skeletal data.
The abstract is short and does not include experimental results or a detailed description of the prototype conditioning. Still, the approach points to a growing interest in using generative models to make motion datasets more compact.