SIH2020_NM381_TeamSkylake

Official code repository of Team Skylake for SIH 2020 grand finale.

FINAL PRESENTATION

https://docs.google.com/presentation/d/1o55oXnzfp19Sy6HjDW_XEgdmVyPkc0L--XAyHjT4BzI/edit#slide=id.p3!

Android app apk

DEMO

DATA SET USED

A large dataset of webcam images annotated with sky regions(90,000) SOURCE: Nathan Jacobs Group

DATA PREPROCESSING

Dataset consists of many corrupted images, so we wrote our own python scripts to remove those corrupted images.

OTHER PREPROCESSING TECHNIQUES

 → Random Rotation
 → Gaussian Blur
 → Normalization

MODELS USED

UNET model with RESNET34 encoder

LOSS FUNCTION USED

Weighted average of Soft Dice Focal Loss

METRIC OF EVALUATION

IOU (Intersection over Union)

TRAINING APPROACH

 → 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

IOU on Validation : → 0.9959

IOU on Test : → 0.9835