My thesis is entitled
Contour and Texture
for Visual Recognition of Object Categories.
See
below
for a video of our new CVPR
2008 prize-winning demo,
Real-Time Object Segmentation with Semantic Texton
Forests.
Contour for Visual Recognition
We as humans
are effortlessly capable of recognising objects from fragments
of image contour. We demonstrated in our
ICCV 2005 paper how an
automatic system can exploit contour as a powerful cue for image
classification and categorical object detection. An
improved multi-scale version of this work
has been accepted for publication in PAMI.
Example object detection
results on the Weizmann horse database.
Green boxes represent correct detections of the horses, red
boxes are false positives, and yellow boxes are false negatives.
The fragments of contour used for detection are visualised in
the final column.
More contour visualisations.
Our technique was applied to a 17 object class
database from TU
Graz. Here are a few examples where the contour fragments
used for detection are superimposed.
Texture for Visual Recognition
A second visual cue is
texture. Our ECCV 2006
paper proposed TextonBoost for simultaneous automatic
object recognition and segmentation, using the repeatable
textural properties of objects. We show how texture,
layout, and textural context can be exploited to achieve
accurate semantic segmentations of images, as
illustrated in the results below and in the
videos available here.
An expanded version has been
accepted to IJCV.
Example
semantic segmentation results.
TextonBoost applied to the MSRC
21 Class Database.
We have recently improved TextonBoost
considerably, making it more accurate and
much faster. This work is summarized in our new
CVPR 2008 paper,
Semantic Texton Forests. Based on randomized
decision forests, our new system is able to run real-time,
illustrated in our demo video:
