+ All Categories
Home > Documents > Learning low-level vision

Learning low-level vision

Date post: 25-Feb-2016
Category:
Upload: aisha
View: 27 times
Download: 0 times
Share this document with a friend
Description:
Computer Examples by Michael Ross. Learning low-level vision. Ising model. Each location has a 50% chance of being 'up' or 'down'. There is a 60% chance that a location has the same value as one of its 8-connected neighbors. - PowerPoint PPT Presentation
17
Learning low-level vision Computer Examples by Michael Ross
Transcript
Page 1: Learning low-level vision

Learning low-level vision

Computer Examples

by Michael Ross

Page 2: Learning low-level vision

Ising model● Each location has a 50% chance of being 'up'

or 'down'.● There is a 60% chance that a location has the

same value as one of its 8-connected neighbors.

● There is an 80% chance that the sensor at a location reports the correct spin.

Page 3: Learning low-level vision

Ising model

True scene. Noise corrupted. Reconstructed.

Page 4: Learning low-level vision

Ising model with Gaussian noise

True scene. Noise corrupted. Reconstructed.

Page 5: Learning low-level vision

Learned optical flow

Page 6: Learning low-level vision

Learned optical flow

Page 7: Learning low-level vision

Learned optical flow

Page 8: Learning low-level vision

Super-resolution

Page 9: Learning low-level vision

Super-resolution

Page 10: Learning low-level vision

Super-resolution

Page 11: Learning low-level vision

Super-resolution

Page 12: Learning low-level vision

Super-resolution

Page 13: Learning low-level vision

Segmentation● An attempt to learn segmentation rules from

examples.● Learn sensor models for each feature.● Construct an MRF with interconnected layers,

one for each feature.● Allow individually insufficient features to

exchange information.

Page 14: Learning low-level vision

Segmentation

Signal: horizontal & verticalgradients.

Scene: edge detected bymotion.

Page 15: Learning low-level vision

Segmentation

...

Page 16: Learning low-level vision

Segmentation

Signal: horizontal & verticalgradients.

Scene: edge detected bybelief propagation.

Page 17: Learning low-level vision

Segmentation● Issues: takes about 25 minutes to produce

result (10 iterations). Why? Considers 100 possible candidates at each location -> ~36 million calculations per iteration.

● Simple features are not very predictive at many locations - better features mean that we need to consider fewer candidates.

● Benefit: learning reduces the number of assumptions and preconceptions.


Recommended