Seeing with Structure: Computer Vision as Inverse Physics

The world is a dynamic 3D environment and if we want computers to reason usefully about the world, they should do so in 3D. To this end, this talk will give an overview of how physics-based priors can be used in computer vision to improve 3d reconstructions from ambiguous videos. It will touch on previous work in part-based modelling [1]; constraints arising in self-supporting structures [2]; the use of Hook’s law in modelling the surface of deformable objects [3]; and the combination of these techniques with shape from shading [4].

[1] Video pop-up: Monocular 3d reconstruction of dynamic scenes (ECCV ‘14)
[2] Part-based modelling of compound scenes from images (CVPR ‘15)
[3] Direct, Dense, and Deformable: Template-Based Non-Rigid 3D Reconstruction from RGB Video (ICCV ‘16)
[4] Better Together: Joint Reasoning for Non-rigid 3D Reconstruction with Specularities and Shading (BMVC ’16)

• Speaker: Chris Russell, Turing Institute
• Thursday 24 November 2016, 14:00–15:00
• Venue: MR14, Centre for Mathematical Sciences.
• Series: CMIH seminar series; organiser: cmih.