CS 143 / Project 2: pb-lite: Boundary Detection / Andrew Ayer <andrew>

Algorithm

The algorithm is implemented as described in the project handout.

1. Create filter bank and masks.
2. Generate a texton map using the filter banks:
   1. Filter the image with each filter in the filter bank.
   2. At every pixel of the image, there is a vector of responses from each of the filters. Use K-Means clustering to cluster these vectors. Currently, I'm using K=64.
   3. Replace each pixel of the image with the cluster ID to create the texton map.
3. Determine how much the texture and brightness is changing at each pixel, by computing texture and brightness gradients.
4. Average the gradient responses and multiply with the sobel/canny baseline to get the final output.

Results

With these enhancements, the F-score is 0.59, which is slightly above the Canny baseline F-score of 0.58, and substantially above the Sobel baseline F-score of 0.45.