kdb
kdb copied to clipboard
cc-ai
Improve flood segmentation model
Improve the current flood segmentation model , see #62
We report below a comparison of a first series of flood segmentation models using mIoU on our hand labeled dataset of 209 images from Keymakr.
DeepLab v2 with a ResNet-101 backbone (current model): - trained on COCO-Stuff 164K dataset - labels water-other, river, sand, sea were merged into the flood category mIoU ~ 79 %
DeepLab v2 with a ResNet-101 backbone + CRF postprocessing: mIoU ~ 80 % Extremely slow (> 10s / image)
Sazara et. al - pretrained VGG-16 + CRF postprocessing: fine tuned on 202 hand-labeled images mIoU ~ 77% slow : 7s/ image (and images had to be resized to have a max size of 600*600 pixels)
Note that the masks for the flooded images will only need to be computed once. Next steps include :
- Fine tune Deeplabv2 model with the flood hand-labeled images
- Qualitative comparison of the models, on our Keymakr test set but also images from other sources
- Report of the failure cases
- Investigate other segmentation models - Deeplabv3, HRNet
