search

amueller / damascene-python-and-matlab-bindings

Python and matlab bindings for the Damascene CUDA implementation of gPB
13 stars 12 forks source link

textons seems to be not right #3

Closed research2010 closed 11 years ago

research2010 commented 11 years ago

Hi Andreas, The textons seems not consistent with the result computed by MATLAB version of CPMC.

This is the original image form VOC dataset. 2007_000032

This is the result by CPMC matlab version. textons_by_cpmcmatlab

This is the result by globalPb_mex. textons_by_gpu_mex

I don't find what's wrong, could you give me some advice, please?

amueller commented 11 years ago

Hey. The textons are the result of clustering, right? So they can not be identified, can they? Cheers, Andy

research2010 commented 11 years ago

Oh, Yes!!! I ignored it. The two results both have values which range 0 to 63. So, they all cluster the features into 64 clusters. Thanks.

amueller commented 11 years ago

you could actually try to compute the rand index between the two (flattened) images, which should be close to one (it's in scikit-learn sklearn.metrics.rand_score) There could be minor differences in the filters and the clustering is randomly initialized and so on, so I'm not sure there is really much of a point to doing this, as long as the results look good ;).