Eye in the Sky

This project was our submission to a challenge in Inter IIT Tech Meet 2018, an annual competition held among the 20 odd IITs in India. With a dataset of just 14 images, we were tasked with segmenting pixels in satellite images into 8 classes - Roads, Buildings, Trees, Grass, Bare Soil, Water, Railways and Swimming pools. The images were multispectral - we had access to a Near InfraRed (NIR) channel along with the usual RGB. Our approach was to handle the task class wise and use a mix of classical computer vision and deep learning. For classes such as Grass, one can in fact use the NIR channel to get an almost pixel-wise accurate segmentation mask without any machine learning! More details can be found in the report here and our code is available here.