Research Project P16:
Learning hierarchical generic parts for object recognition

Research Project Goal

Implement a system for unsupervised learning of hierarchical generic parts from natural images.

One of the commonly used approaches for object recognition assumes representations of objects in terms of parts and spatial relationship between them. Various methods for detection and learning of generic parts have been recently proposed, assuming as a front end commonly used feature detectors. The goal of this project is to explore an approach for learning hierarchical genetic part based models, where constituent parts are characteristic of object shape (e.g. pieces of contours, junctions, intersections).

Sample Data

Any of the widely available databases used for object recognition COIL database, Pascal database, Caltech 101 database. As a subset of object categories, consider first those whose visual shape is 'intuitively' discriminative (e.g. cows vs airplanes)

Tasks

Given a set of contours detected from the training images, sample part candidates of a particular size. Build visual vocabulary of elementary part (or suggest other statistical means how to select the sparse set of candidates). Based on non-coincidental co-occurrences of certain neighboring parts, learn composite parts at higher level of visual hierarchy.

Issues

• choice of the metric of how to compare individual parts (i.e. what does it mean for the parts to be the same)
• how to robustly determine the composite parts and how to represent their relationships so as to achieve invariance to scale and viewpoint change (rotation, translation).

Research Project Status

student names here

Point of Contact

Jana Kosecka

Midterm Report

not yet submitted

Final Report

not yet submitted