search

Neuroglycerin / neukrill-net-tools

Tools coded as part of the NDSB competition.
MIT License
0 stars 0 forks source link

Cropping Augmentation #22

Closed gngdb closed 9 years ago

gngdb commented 9 years ago

As in the Baidu paper, random cropping within the image.

gngdb commented 9 years ago

This might be more important than I originally thought. The images are not perfectly reliably cropped; some plankton overlap the edges. Found an example by accident: train/hydromedusae_sideview_big/100058.jpg

When working on this, probably worth going through the images finding those that overlap to get an idea of how much random crop to apply.

scottclowe commented 9 years ago

We should make sure we DON'T crop any members of the tunicate class, because there is a subclass called "tunicate_partial" where all the members look like cropped versions of the "tunicate salp" class. If we were to crop the tunicate salp training data, it would only lead to mislabelling with the tunicate partial class. Similarly, we also need to be careful with respect to the "hydromedusae_partial_dark" and "siphonophore_partial" classes.