3D-based Object Understanding in The Wild
Event details
Date | 04.06.2018 |
Hour | 09:00 › 11:00 |
Speaker | Edoardo Remelli |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Dr. François Fleuret
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Martin Jaggi
Abstract
Recovering the 3D shape of an object from a single perspective has always been central to Computer Vision. Despite recent advances, current state-of-the-art single view 3D mesh reconstruction techniques either do not perform well on real data or rely on strong manual input. In this research proposal we first review three different approaches to 3D reconstruction from monocular imagery [1], [2], [3]. Then, building upon them, we propose a framework generating 3D meshes from RGB input which, making use of recurrent neural networks, can generate meshes with arbitrary topologies. We envision the first method able to reconstruct arbitrary objects given monocular sensor data in unconstrained settings, and we are confident this will be a milestone towards general 3D-based object understanding.
Background papers
What shape are dolphins? Building 3D morphable models from 2D images, by Cashman T.J., Fitzgibbon A.,W.,
Learning category-specific mesh reconstruction from image collections, by Kanazawa A., et al.
3D-PRNN: Generating shape primitives with recurrent neural networks, by Zou, C., et al.
Exam president: Dr. François Fleuret
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Martin Jaggi
Abstract
Recovering the 3D shape of an object from a single perspective has always been central to Computer Vision. Despite recent advances, current state-of-the-art single view 3D mesh reconstruction techniques either do not perform well on real data or rely on strong manual input. In this research proposal we first review three different approaches to 3D reconstruction from monocular imagery [1], [2], [3]. Then, building upon them, we propose a framework generating 3D meshes from RGB input which, making use of recurrent neural networks, can generate meshes with arbitrary topologies. We envision the first method able to reconstruct arbitrary objects given monocular sensor data in unconstrained settings, and we are confident this will be a milestone towards general 3D-based object understanding.
Background papers
What shape are dolphins? Building 3D morphable models from 2D images, by Cashman T.J., Fitzgibbon A.,W.,
Learning category-specific mesh reconstruction from image collections, by Kanazawa A., et al.
3D-PRNN: Generating shape primitives with recurrent neural networks, by Zou, C., et al.
Practical information
- General public
- Free
Contact
- EDIC - [email protected]