Project2 : pb-lite Boundary Detection

Jihoon Lee

Overview

The goal of this project is to develop a pb-lite boundary detection algorithm which is introduced in the recent work of Arbelaez, Maire, Forwlkes, and Malik. TPAMI 2011(pdf). The simplified version of pb uses canny and sobel detection algorithm as a baseline.

Approach

This simplified pb-lite boundary detection uses the combination of baseline edges detected by sobel/canny and thresholded texture gradient + brightness gradient. The gradients are computed by measuring chi-square distances for the performance.


The detail algorithm is by following:
  1. Compute baseline(Canny, Sobel) edges
  2. Generate Filter bank and Half-Disc Masks
  3. Build a texton map using filter bank and kmean
  4. Compute texture gradient and brightness gradient using chi-square distances
  5. Combine gradients and baseline edges to obtain final edges

Result

In order to achieve better performance than baseline detectors, I multiplied the mean of gradients to canny edge detector. The F-score shows that the my version of pb-lite is slightly better than both baseline 1 and 2 edge detector.

baseline1(sobel) baseline2(canny) pb-lite
ODS F(0.38,0.55) = 0.45 [th = 0.09] F(0.66,0.51) = 0.58 [th = 0.15] F(0.69,0.51) = 0.59 [th = 0.11]
OIS F(0.38,0.55) = 0.45 F(0.70,0.50) = 0.59 F(0.68,0.52) = 0.59
Area_PR 0.21 0.50 0.51

Filter Bank and Half-disc Masks

A collection of oriented derivative of Gaussian filter

The half-circle masks

Edge Images