CVPR Daily - Tuesday

DAILY T u e s d a y Learn3dgen 25 3D mesh reconstructions of birds from images (from Jitendra’s talk, Kanawaza et al, ECCV 2018) Later in the day, Sanja Fidler presented how synthetic 3D data is useful for training AI agents and vision algorithms in self-driving cars, indoor robots, and medical imaging. As an example of the challenges of learning 3D generative models, let’s look at the problem of generating 3D shapes. Paul Guerrero , research scientist at Adobe , presented exciting joint work between Adobe , Stanford and UCSD on training neural networks to encode the 3D structure of a shape in a shape embedding space to allow for generating new shapes and editing 3D shapes. Their work uses the PartNet dataset (see Figure), which provides annotated object part hierarchies on top of ShapeNet, a collection of common 3D objects. The methods that Paul presented are capable of encoding a 3D shape and its hierarchical structure (in terms of parts such as the seat and armrests of a chair) into a latent space from which the shape and its structure can be reconstructed (see Figure). Visualization of PartNet dataset 3D objects and their part annotations (from Mo et al, CVPR 2019).