This software package provides tools to extract contour-based mid-level features, and to extract contour segmentations from images. This tool is highly efficient in speed while maintains high accuracy in contour detection. Also, [1] shows that extracted mid-level features provide additional information for object and pedestrian detections.

1. Download Piotr's Image & Video Matlab Toolbox (http://vision.ucsd.edu/~pdollar/toolbox/doc/)
   SketchTokens/toolbox/ should have channels, classify, filters, images, matlab, etc,.
2. Download Berkeley Segmentation Data Set and Benchmarks 500 (http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/resources.html)
   SketchTokens/data/BSR/ should have BSDS500, bench, and documentation.
3. Pre-trained models can be downloaded from: http://people.csail.mit.edu/lim/lzd_cvpr2013/st_data.tgz
4. Look up stDemo.m for how to train and test our code

Please cite the following paper if you end up using the code:
[1] Joseph J. Lim, C. Lawrence Zitnick, and Piotr Dollar. "Sketch Tokens: A Learned Mid-level Representation for Contour and and Object Detection," CVPR2013.

Copyright 2013 Joseph Lim [[email protected]]

Please email me if you find bugs, or have suggestions or questions!

Licensed under the Simplified BSD License [see bsd.txt]

Note: There is a patent pending on the ideas presented in this work so this code should only be used for academic purposes.