EE Seminar: Modeling and Learning Similarity of Shapes, Images and Signals
Speaker: Roee Litman
Ph.D. student under the supervision of Prof. Alex Bronstein
Monday, March 27th, 2017 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Modeling and Learning Similarity of Shapes, Images and Signals
Abstract
More than a decade ago, a major part of computer vision research was dedicated to engineering and designing the best way to capture meaningful features in images. So was the case in geometry processing and shape analysis. With time, the size of annotated datasets grew, also becoming more realistic and challenging. Together with the increase of low-cost computational power, this turned the focus of research more and more towards learning those features from the data themselves. While problem modeling is still a crucial part of research, learning-based methods are particularly successful whenever noise-invariance is harder to model, especially when the data are deviating from theory.
I will present several new methods and advances for the problem of measuring similarity and establishing correspondence between shapes, images and signals. The topics covered here progress from `designed' to `learned' in a gradual manner. First, there are some cases where the `right' model can solve the problem in a manner close to optimal, as shown for the problem of shape correspondence. Next, a model can be designed such that some small parts are allowed to adjust according to examples in order to improve performance, as shown in the case of shape descriptors. Finally, in some cases the model makes very mild assumptions and is almost completely learned from examples, as in the case of task specific sparse models.
