search

biigle / maia

:m: BIIGLE module for the Machine Learning Assisted Image Annotation method
GNU General Public License v3.0
2 stars 3 forks source link

Sliding window processing of large images #92

Open mzur opened 2 years ago

mzur commented 2 years ago

MAIA disallows processing of images that are too large to fit into GPU memory. We could implement processing in a sliding window fashion (on smaller image crops, e.g. 4096x4096) if the image is too large. This could allow processing of tiled images or otherwise too large images and make #91 unnecessary.

Ideally, the sliding window should have an overlap so objects are not cut in half and missed. How to merge the detections in the overlap regions? Maybe a greedy method where any two detections with more than 25% (50%?) overlap are merged. In case of many overlapping annotations, those with the largest overlap are preferred.

This could be implemented entirely in Python, using pyvips to cut the images.

mzur commented 2 years ago

Could novelty detection be performed in a sliding window fashion, too? Otherwise, disable novelty detection if the volume contains too large images (see #91).