Eff-UNet: A Novel Architecture for Semantic Segmentation in Unstructured Environment

Abstract

Since the last few decades, the number of road causalities has seen continuous growth across the globe. Nowa-days intelligent transportation systems are being developed to enable safe and relaxed driving and scene understanding of the surrounding environment is an integral part of it. While several approaches are being developed for semantic scene segmentation based on deep learning and Convolutional Neural Network (CNN), these approaches assume well structured road infrastructure and driving environment. We focus our work on recent India Driving Lite Dataset (IDD), which contains data from unstructured driving environment and was hosted as an online challenge in NCVPRIPG 2019. We propose a novel architecture named as Eff-UNet which combines the effectiveness of compound scaled EfficientNet as the encoder for feature extraction with UNet decoder for reconstructing the fine-grained segmentation map. High level feature information as well as low level spatial information useful for precise segmentation are combined. The proposed architecture achieved 0.7376 and 0.6276 mean Intersection over Union (mIoU) on validation and test dataset respectively and won first prize in IDD lite segmentation challenge outperforming other approaches in the literature.