ebgoldstein / iou_maps Goto Github PK

100 images will be labeled by 5 people. Need to do IRR metrics using Krippendorf's alpha (ordinal)..

calculate mIoU, per class IoU, and Dice for each image, and record results

merge results with other info (lat, lon, storm, date, etc.)

create maps with resulting metrics data

• label with a Likert scale, using Sniffer

compare labels with predicted IoU, and make ridge plot

entropy, PHOG/HOG, FFT, saliency, colorfulness, lossy compression

• pick 1000 images not from training data
• make prediction (full UNet)
• make image + overlay pairs

cut off decoder, and add Dense layers after the bottleneck to predict metrics (IoU, Dice, etc.)

predict all images with the metrics model, merge results to large dataframe