The goal of this assignment is to develop a simplified version of the pb boundary detetor by Arbelaez, Maire, Fowlkes, and Malik. TPAMI 2011 (pdf) to beat the baseline boundary detector.
Algorithm
The algorithm used in practice is simplified into four steps as below:
- The first step is to build filter bank and half-disc masks for the following use, which need only to be defined once and will be used on all images. The oriented derivatives of Gaussian filters built is as follows:
The half-disc masks are like:
- The Second step is to generate the texton map of an image by clustering the filtered image with multiple responses using kmeans algorithm. The result is a single channel image which we could visualize as:
Different colors indicate pixels belong to different classes.
- Next is to compute texture gradient and brightness gradient by building hitograms over regions of interest and comparing distributions in left/right half-disc pairs at each pixel.
- The final step is to combine information from the features with a baseline method.
Richer Filters
We will try to use more orientation selective filters and center-surround filters. The simple oriented box filter is like:
One of the center-surround filter is the Laplacian of Gaussian:
Another is the Difference of Gaussian filter which is usually used as the approximation to LoG:
Richer Masks
The square masks could also be built to be compared to the circular masks:
Richer features
Besides tg and bg, the three color channels are also added as features in the code.
Results
The result ROC curve is:
And the F score reported is:
F(0.65, 0.56) = 0.60
which could beat the baseline1 and baselin2.