Computer Vision News - July 2020

Despoina Paschalidou 53 are more semantic and parsimonious . These are some of the attributes that allow for the argument that deep learning networks are able to learn an interpretable representation, when we tend to think that neural networks only learn how to generalize. Looking back, Despoina recalls the project very nearly didn’t happen at all, as her initial ideas and attempts didn’t work out as well as she’d hoped. She almost gave up hope, but after a moment of clarity, she began designing the multi-level representations that she is proposing here. Although people haveproposedsolutions to thisproblem before, the majority have tried to solve it in a supervised way with supervision in the structure of the object. This method doesn’t have any supervision, either on the primitive parameters, nor on the structure of the object . She believes it is the first work to do this in an unsupervised fashion. Despoina’s supervisors on this work are Andreas Geiger at the Max Planck Institute, and Luc van Gool at ETH Zürich . “Both have inspired me,” she says. “They have consolidated these ideas and my vision towards what I want to do next.” As has Ali Osman Ulusoy , co-author on her 2018 and 2019 CVPR submissions: “His help and guidance has helped me a lot.” She says it’s easy for PhD students to get excited about the details of their work and miss the big picture of how it will be useful in the real world. “Primitive-based representations have these special properties that they are very semantic,” she explains. “This is a property that is useful if we want to be able to capture high-level content in a scene. Not just – this is a chair; this is a table. I think my work can be used in applications such as AR and VR. One “One application that I would love to see happen before I finish my PhD is to scale this work into real 3D scenarios and model interactions between parts in a scene.” Best of CVPR 2020