For segmenting images with many small not-spacially related objects, it is useful to split training images into smaller examples. A sliding-window network can then be applied to process the images. Or, in a fully convolutional case with no dense layers, networks which learn on small images could be applied to large images directly.
This module should take a training image + mask pair and a snipping step size (allowing overlapping training examples), and output a larger set of training images for concatenation into a common training dataset.
For segmenting images with many small not-spacially related objects, it is useful to split training images into smaller examples. A sliding-window network can then be applied to process the images. Or, in a fully convolutional case with no dense layers, networks which learn on small images could be applied to large images directly.
This module should take a training image + mask pair and a snipping step size (allowing overlapping training examples), and output a larger set of training images for concatenation into a common training dataset.
https://github.com/bdzyubak/Bogdans-Deep-Learning-Sandbox/issues/5