Seminar Announcement

Minglun Gong

Department of Computer Science
Memorial University of Newfoundland

What Do Images Tell Us about the Geometry of an Object?

Department of Computer Science
Thursday, April 9, 2015, 1:00 p.m., Room EN-2022

---

 

Abstract:

An image worth a thousand words. From images, we humans are able to infer the 3D shape of an object and to decompose the object into semantically meaningful parts. Now, is it possible to teach computers to do these tasks?

Two recent research projects that work along this direction will be presented in this talk. The first one investigates how the 3D modeling of flower head can be facilitated using a single photo of the flower. The core idea is that flower head typically consists of petals of similar 3D geometries, yet their observed shapes on 2D images vary due to differences in projecting directions. Exploiting this variation allows us to reconstruct the 3D geometry of the petals from a single image.

The second project studies how to segment 3D models into semantically meaningful parts based on knowledge learned from labeled 2D images. Here the input 3D model is treated as a collection of 2D projections, which are labeled using training images of similar objects. The 3D model is then segmented by summarizing the labeling for its projections. Here the key is, for each query projection, how to retrieve objects with similar semantic parts and transfer their labels over.