Add example using non-georeference imagery

[lewfish]

Medical would be good? Single band would also be good to test. Something with a permissive license.

[simonkassel]

is this a high priority? I could potentially take it on. Do we have any ideas about what we would want to use?

Here are a few options I found from kaggle:

Pneumonia detection challenge

• object detection, locate lung opacities in chest radiographs
• labels come in a csv of long/lat bounding boxes
• images are in DICOM format which seems common for medical stuff

CT medical images

• just a dataset, not a competition
• seems like it's just a dataset of images with age and some boolean field called contrast (not sure exactly what it signifies)
• we could potentially classify images by either of these categories (although age is continuous)
• DIRCOM and TIFF format
• looks like these images are only 512x512

NIH chest x-rays

• just dataset, not a competition
• 112,120 total images with size 1024 x 1024 (not sure if this is big enough to make it useful to use RV)
• seems like this would be classifying images into 15 categories (14 diseases + no disease category)
• looks like there are a limited number of labeled bounding boxes for specific locations of disease within imagery

also here is a curated list of machine learning medical imagery datasets. I haven't looked through them yet but it looks like a lot require registration

[simonkassel]

looks like the pneumonia detection images are only 1024x1024 as well

[lewfish]

Rob is working on https://camelyon16.grand-challenge.org/ so I would hold off on this.

[simonkassel]

ok, right, I forgot about that