search

RockStarCoders / alienMarkovNetworks

Using MRFs and CRFs for computer vision problems.
21 stars 9 forks source link

investigate region neighbour co-occurrences #19

Closed jsherrah closed 10 years ago

jsherrah commented 11 years ago

it might be estimated from few examples due to void pixels.

Use the matlab script example to 'fill the gaps'?

jsherrah commented 11 years ago

void is only a big issue at pixel level. At superpixel level hopefully a region will be assigned a non-void class.

Need to investigate the stats of how often that occurs.

jsherrah commented 11 years ago

add extra weight param to split out degree and adjacency prob.

jsherrah commented 11 years ago

Added a stats trace-write to the createFeatures code:

**Processed total of 591 images 196036 out of 589933 adjacencies were ignored due to void (33.23 %)

1/3 of adjacencies are ignored due to void. I think something needs to be done, will investigate further what the reasons are...

jsherrah commented 11 years ago

Note this doesn't hold up the classification work, only MRF inference.

jsherrah commented 10 years ago

On further consideration I don't think this is a good approach. A sizeable object consists of lots of super-pixels, and they are for the most part adjacent to regions of the same class. Therefore the co-occurrences will be quite weak. On top of that we have difficulties getting training data so that noise will have a large effect.

TextonBoost uses local neighbouring rectangles to encapsulate information about local spatial relationships anyway.

There is other work on the co-occurrence of labels in the same image, independent of their frequency or location.

I'm going to stop working on this, and perhaps even remove it from the code so nobody uses it by mistake.