create multi-class scene labelling example

[jsherrah]

create file sceneLabelN.py to do multi-class scene labels with first-order potentials.

The plan is to reproduce the work of Shotton et al, ECCV 2006:

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.83.519

Probably use stair vision library, but needs investigation.

[jsherrah]

Having read that paper, it's rooly complicated. Since the emphasis at the mo is on the CRFs, I will use a simpler approach. The unary potentials will be some classifier or probabilistic model of the local texture per class. We can start with something really simple, like HOG and/or colour histograms, and then add complexity until finally reproducing state of the art (which is unlikely to be a 2006 paper).

Shotton 2006 uses Boykov and Jolly 2001 for inference. B&J use a new max flow method from their comparison paper. This is the Boykov, Kolmogorov C implementation that seems to get used everywhere (ALE, PyMaxflow).

I have just realised PyMaxflow can do multi-class labelling and alpha expansion using fastmin:

maxflow.fastmin.aexpansion_grid(D, V, max_cycles=None, labels=None)

Let's continue to use PyMaxflow.

[amb-enthusiast]

Created a 3-class example (foreground "midground" and background). We could close this issue, noting that the example model is simple.

PyMaxflow's alpha expansion and alpha-beta swap methods accept pairwise potentials between class-labels - not between label variables/nodes.

Probably need to move away from PyMaxflow long term. May need other approximate inference techniques (belief propagation, MCMC).

[jsherrah]

let's close it once we have a labelling on the ms data set.

[amb-enthusiast]

Hey J, how do you feel about closing this issue now? We've got N-class working on the MSRC dataset now, and even have superpixels.

[jsherrah]

Ok

Sent from my iPad

On Aug 19, 2013, at 8:08 AM, Ant B notifications@github.com wrote:

Hey J, how do you feel about closing this issue now? We've got N-class working on the MSRC dataset now, and even have superpixels.

— Reply to this email directly or view it on GitHub.