scottclowe
commented
9 years ago For the implementation, use
neukrill_net.utils.Settings(settings.json)
and load images with
neukrill_net.utils.load_rawdata.
and loop until an image is no good.
(See train_bow.py in work repo for example of how to load data with this function.)
Alternatively, we could write a processing wrapper function which is applied to each image on load and returns True or False for each image. That would probably be better. Although it wouldn't be able to halt at the first False.
We need a function which tests a keypoint detector by trying it on all the data (train and test) and checking it finds an appropriate number of keypoints for all of them.