Introduction
AND-OR Template (AOT) is a probabilistic image model for object recognition. It has the following characteristics:
- Easy to train: The AOT model can be learned/trained from as few as 10 positive example images. No bounding boxes labeling needed. No negative examples needed. The training finishes within minutes.
- Compositional (AND): The 2-level object-part-primitive hierarchy makes AOT a great model for articulated objects. The AND means composition of constituent parts into a larger part.
- Deformable (OR): The constituent parts are allowed to transform locally around their canonical locations within their parent part. OR means deformation, and different ways to compose smaller parts into a larger one.
- Generative: The AOT defines a probability density function supported on image pixels (NOT features values). Thus it enables accurate template matching down to pixel level. It is also easy to visualize the object template and the match result.
AND-OR Template for Egret (automatically learned) |
|
Detection of object, parts and primitives |
Source code: training, detection.
More examples and comparison with latent SVM.
Publication
Learning AND-OR Templates (AOT) for object recognition and detection
PDF
Zhangzhang Si and Song-Chun Zhu
IEEE Transactions on Pattern ananlysis and machine intelligence, (accepted) 2013.
Zhangzhang Si and Song-Chun Zhu
IEEE Transactions on Pattern ananlysis and machine intelligence, (accepted) 2013.
Learning active basis model for object detection and recognition
PDF
Ying Nian Wu, Zhangzhang Si, Haifeng Gong, and Song-Chun Zhu
International Journal of Computer Vision, vol.90, no.2, pp 198-235, 2010.
Ying Nian Wu, Zhangzhang Si, Haifeng Gong, and Song-Chun Zhu
International Journal of Computer Vision, vol.90, no.2, pp 198-235, 2010.
Learning Hybrid image Template (HiT) by Information Projection
PDF
Zhangzhang Si and Song-Chun Zhu
IEEE Transactions on Pattern ananlysis and machine intelligence, vol.34, no.7, pp 1354-1367, 2012.
Zhangzhang Si and Song-Chun Zhu
IEEE Transactions on Pattern ananlysis and machine intelligence, vol.34, no.7, pp 1354-1367, 2012.