Official code repository of Team Skylake for SIH 2020 grand finale.
https://docs.google.com/presentation/d/1o55oXnzfp19Sy6HjDW_XEgdmVyPkc0L--XAyHjT4BzI/edit#slide=id.p3!
A large dataset of webcam images annotated with sky regions(90,000) SOURCE: Nathan Jacobs Group
Dataset consists of many corrupted images, so we wrote our own python scripts to remove those corrupted images.
→ Random Rotation
→ Gaussian Blur
→ Normalization
UNET model with RESNET34 encoder
Weighted average of Soft Dice Focal Loss
IOU (Intersection over Union)
→ used model pre trained on IMAGENET.
→ progressive training due to huge size of data set
→ started with 20 percent of dataset to provide warm start for training
→ Gradually increased to 70 percent in step of 5
→ this was used along with 5 fold cross validation due to the lack of diversity in images
→ tested on 30 percent images