Weakly supervised learning of hierarhical active basis

Code and data: (ZIP).	Change Log
Run StartFromHere.m in matlab to learn the hierarchical active basis template. Run GenerateHtml.m to generate this webpage.

Training examples

We learn a 2-level hierarchical active basis template from images with articulated objects with unknown scale, rotation and translation. The object template size is 180 * 180 pixels; the part template size is 60 * 60 pixels. Each part may translate and rotate locally relative to the object center. And each edge element can translate and rotate locally relative to the part center. We use EM-type algorithm to obtain the part and object templates as well as the hierarchical deformation on each image. Number of EM iterations = 5. Each part is an active basis model of Gabor elements with local shift <= 3 pixels, and orientation shift <= 1 * PI/16. In total, there are 120 Gabor elements in the object template.

Learning from example 1

The template is initialized from the following example image.

Learned object template

Shape matching by Hierarchical Active Basis

A hierarchical active basis (HAB) template is composed of a small number of partial templates which can translate and rotate locally relative to other partial templates.The following pairs of images show the matched HAB templates on example images.