GitLab

Specifically, about what the compatibility and kernel parameters are, about missing the weights, and some pointers about learning.

There was a sudden interest in this, both via issues (#23) and direct e-mail almost simultaneously. Odd 😄

Clearer and more unified unary utilities.

Examples fixFixes to the examples, it's only one now, but it's better!

Fix off-by-one error in `compute_unary` utility function.

Due to the special case of '0' indicating no label and being assigned uniform probs, and M being the number of labels not including this I believe the original function was not correct, as the highest label would be out of bounds, and the first label mapped to row 1 of the unary matrix.

The above minor change should fix this, ergo label 1 corresponds to row 1, the number of rows is equal to M, labels go from 1 to M.

Fixed inference example (wrong keyword argument)

Fixes #13 continuation.

So it can be used usefully without matplotlib.

fixed a bug in create_pair_gaussian method in utils.py

Especially, installs `__init__.py` which is necessary for the import.

Util functions

Fix height width order in dense_inference.py example

Added several useful util functions to create features. This functions are not limited to the simple 2D case.

This has to do with C-ordering versus Fortran ordering, also known as https://en.wikipedia.org/wiki/Row-major_order

This bug appears with the following parameters:

d.addPairwiseGaussian(sxy=3, compat=0)
d.addPairwiseBilateral(sxy=3, srgb=5, rgbim=img, compat=50)

And this is confirmed by this other Python wrapper for densecrf:

see the order (D, W, H) at this line:
https://github.com/mbickel/DenseInferenceWrapper/blob/master/denseinference/lib/libDenseCRF/densecrf.cpp#L139

while the original 2D code is (H,W).

Example dense inference

adding Python example for dense inference adapted from the original C++ example (http://www.philkr.net/home/densecrf)