Learning Active Basis Models by EM-type Algorithms
Zhangzhang Si, Haifeng Gong, Song-Chun Zhu, and Ying Nian Wu
This is a review paper we were invited to contribute to a special issue on EM
algorithm. The paper uses more
mathematically explicit notation, and introduces the model and leraning by
gradually developing the shared matching pursuit algorithm.
Important: We have stopped updating this page.
Reprint |latex files
- In the active basis model, both the perturbations and the coefficients of the
basis elements are hidden variables that seek to explain each individual training
- We can add more hidden variables such as unknown categories, poses, locations,
scales and orientations of the objects. The learning with such hidden variables is
to be accomplished by EM-like algorithms. The idea is to seek better alignments by
allowing more flexibilities.
The reproducibility page for IJCV paper (Section 2) has the most updated
source code. The code below is only for reproducing the
We have added many new results not presented in the paper. In particular, all the results in
Experiment 4 are new to the paper.
Research reported on this page was supported by
NSF DMS 0707055 and NSF IIS 0713652.
Back to active basis homepage