Unsupervised Texture
Segmentation on Graphics
Hardware
Chris White
Data Driven Graphics
Problem Statement
Given an image, segment into regions
of similar texture
Perform segmentation quickly –
hopefully in real time
Method
Transform each pixel in the input image
into a feature vector that captures local
structure
Cluster resulting feature vectors into
Feature vectors
Smooth the structure tensor via nonlinear diffusion
GPU implementation is straight forward
Segmentation of Features
K-means or Level set
Non-trivial GPGPU work
Metrics for Success
Efficiency – wall clock comparison to CPU
algorithm
Accuracy – for synthetic scenes consisting of
objects textured with examples from Brodatz
texture database, ground truth is known and
error can be computed
Several Brodatz textures