Our submission for the AutoNUE: Scene Understanding Challenge 2018 (ECCV’18), won the intel travel-grant. For more information, please look at the research works section!

H2 Abstract:

Semantic Segmentation is much more challenging in the presence of multiple similar classes, and high intra-class variations. Datasets such as AutoNUE model real-life scenarios, and feature. Large intra-class appearance variations, Presence of low-shot or novel classes. In such scenarios, simple deep-learning approaches can have high confusion among similar classes, and hence perform poorly. To yield improved performance in such a unconstrained dataset, it is important to clearly discern the differences between confusing classes. Hence, in our approach, we propose a novel Expertise-Layer to enhance the learned model’s discerning ability.