Real-time Semantic Segmentation with Context Aggregation Network

With the increasing demand of autonomous systems, pixelwise semantic segmentation for visual scene understanding needs to be not only accurate but also efficient potential real-time applications. In this paper, we propose Context Aggregation Network, a dual branch convolutional neural network, with significantly lower computational costs as compared state-of-the-art, while maintaining competitive prediction accuracy. Building upon existing architectures high-speed segmentation, design high resolution effective spatial detailing and context light-weight versions global aggregation local distribution blocks, potent capture both long-range contextual dependencies required low overheads. We evaluate our method on two datasets, namely Cityscapes dataset UAVid dataset. For test set, model achieves state-of-the-art results mIOU 75.9%, at 76 FPS an NVIDIA RTX 2080Ti 8 Jetson Xavier NX. regards dataset, proposed network score 63.5% execution speed (15 FPS).