Semantic Segmentation Research

First Project: ResUnet-CRF (RUF) for Sea Ice Floes

This project, done in collaboration with Anmol Nagi and Devinder Kumar, aimed to segment sea ice floes (floating ice) and was recently submitted for publication. In this project we performed semantic segmentation on satellite imagery from RADARSAT-2 of Hudson Strait in the Canadian Arctic in order to assist ships in navigating the Arctic.

We proposed a segmentation model tailored for detecting ice floes in SAR images which exploits the advantages of both convolutional neural networks and convolutional conditional random field (Conv-CRF) in a combined manner. A residual UNET (RES-UNET) computes expressive features to generate coarse segmentation maps while a Conv-CRF exploits the spatial co-occurrence pairwise potentials along with the RES-UNET segmentation maps to generate final predictions. The whole pipeline is trained end-to-end using a dual loss and outperforms conventional segmentation networks such as UNET, DeepLabV3, and FCN-8.

Second Project: Semi-Supervised Segmentation of Drone Images and Video

This project is in its beginning stages and involves segmenting various types of ice from a large drone data set with few labels. Weakly supervised methods such as point supervision are being explored as well as adversarial learning which uses a GAN-like architecture in a fully convolutional manner to differentiate the predicted probability maps from a ground truth segmentation distribution.