An opposition to LOCUS

1
An opposition to LOCUS

• David Lee

2

• Unsupervised?
• Poor dataset
• and
• How to take advantage of it with top-down
  generative approach

3
(No Transcript)
4
Unsupervised?

• Images themselves are already a crude
  segmentation
• Assumptions (Tell the system that)
• There exists one horse,
• facing left,
• occupying 15-30 of the image,
• centered (although authors claim theyre not),
• strong edge between foreground background
• small color/texture variance within foreground
• ? Weizmann horses Caltech 101 !!

5
Mean of Canny edges
6
Mean of hand-labeled mask
7
Top-down generative approach

• When the data has strong constraints,take
  advantage of that high level prior knowledge.
• If you know what the data is like, let the
  generative model be it.

8
My Universal Horse Segmentation
MUHS
70
9
(No Transcript)
10
My Universal Horse Segmentation
0.536240 0.880552 0.493996 0.708324 0.718089
0.530342 0.802334 0.545985 0.470904