search

openearth / flamingo

Python toolbox for coastal image analysis
http://flamingo-image.readthedocs.io/
GNU General Public License v2.0
9 stars 1 forks source link

Fix compatibility problem with channels #6

Open hoonhout opened 9 years ago

hoonhout commented 9 years ago

The channels extracted for the argusnl dataset and the paper are not compatible with the current toolbox.

In August 2014 gabor channels were extracted varying over both frequencies (3) and sigmas (3). The maximum response obtained over the thetas (4) was used as channel value. Besides, gaussian channels were extracted (3). In total we had 3x3+3 = 12 channels added to the 4 intrinsic channels RGB + grayscale. The case studies also used this configuration.

In December 2014, we removed the 3 sigma iterations from the gabor channels since they are redundant with the frequencies. We therefore now have 3+3 = 6 channels added to the 4 intrinsic ones. These channels are initially extracted for the zm_lagoon dataset as well.

There is a bug in the code in channels.get_number_channels since it assumes that there will be # frequencies x # theta gabor channels (3x4 = 12), while in fact only the maximum theta is used, so there will be only # frequencies gabor channels. This should be fixed.

After that the whole preprocessing and training of the argusnl dataset should be redone in order to make the dataset compatible with the current toolbox. At this stage we deviate from the paper. This will take a week of computational time.

In the mean time Sierd cannot continue on the zm_lagoon dataset. So, it might be useful to restore the original code where gabor channels are iterated over the sigmas as well and then preprocess the zm_lagoon dataset to make it compatible with the current argusnl model.

hoonhout commented 9 years ago

Started branch "paper" to restore the toolbox as it was during this study