Real Virtual Humans (Talk)

With the explosive growth of available training data, 3D human pose and shape estimation is ahead of a transition to a data-centric paradigm. To leverage data scale, we need flexible models trainable from heterogeneous data sources. To this end, our latest work, Neural Localizer Fields, seamlessly unifies different human pose and shape-related tasks and datasets though the ability - both at training and test time - to query any arbitrary point of the human volume, and obtain its estimated location in 3D, based on a single RGB image. We achieve this by learning a continuous neural field of body point localizer functions, each of which is a differently parameterized 3D heatmap-based convolutional point localizer. This way, we can naturally exploit differently annotated data sources including parametric mesh, 2D/3D skeleton and dense pose, without having to explicitly convert between them, and thereby train large-scale 3D human mesh and skeleton estimation models that outperform the state-of-the-art on several public benchmarks including 3DPW, EMDB and SSP-3D by a considerable margin.