Poster T47, Oral Tues. 11:30 Scene Segmentation with CRFs Jakob Verb eek (INRIA) Learned from Partially Lab eled Images Bill Triggs (CNRS)

· Conditional Random Field infers region-level lab eling of image content
l S

h

ocal features and large scale aggregates IFT and color features

x x x

x x

x x

x

· CRF learning from partial lab eling
verage over possible label completions using Loopy BP to approximate Free Energy contrast automatically infers hidden lab el boundaries
a
y

y y

y y

y y

y

CRF model with aggregate features

· Exp erimental Results
SRC, Sowerby, Corel datasets: 7-9 classes, 100-240 images 40% lab eling is almost as go o d as 70% the most informative aggregates are the image-wide ones
M

Some typical image segmentations