Hugin++ Files

Celeste Training:

1. Fire up the get_flickr_clouds.pl script and download some images from Flickr.
Edit the scipt to fill in your own authentication token (you'll need to be logged
into Flickr). Change the search terms and the directory paths.

2. Edit your training images and put them in 

cropped_images/

and

masked_images/

In cropped images, files named _______P.jpg are positive examples (e.g. contain
nothing but cloud), and files named _______N.jpg are negative examples (e.g. contain
no cloud). Images placed in masked_images must have a correpsonding mask file with
the same name but a M.jpg extension. In the mask, clouds must be black and non-cloud
white.

You will need at least 20 images. Celeste was trained on about 170.


3. Run ./create_svm_data_10_fold.pl

This runs ./celeste_train on each of the images, and generates SVM vectors which it
puts in image_data/ . It also creates the 10 folds used for 10-fold cross validation.

4. Run ./svm_learn_class_10_fold.pl

You'll need to set the paths for the svm_light files at the top of the script unless
they are in your path. This step will take a while to learn and classify (best run on
a cluster..). Results are written to files in the results/ directory.

5. Run ./parse_results.pl to show the results for each run, using MCC as the scoring
metric, and the number of folds complete in brackets.

6  Adjust parameters string in create_svm_data_10_fold.pl (and possibly in the 
CelesteGlobals.cpp file, in which case recompile then goto step 3).

7. Go back to step 4.


Contact me @ timnugent@gmail.com if you need any help.