Supervisors: Dr. Shuhang Gu, Dr. Radu Timofte
Semantic segmentation requires good spatial information and large receptive field for a promising segmentation result. And with the rise of machine learning with assistance of GPU, convolutional neural network (CNN) has been proven as an effective way to achieve this task. At the start, most of the work relates to enlarge receptive field by making neural network go deep. However, this method sacrifices spatial information for large scale of receptive field. Performance of this method is, although somehow optimistic, quite slow. Later work in semantic segmentation focused on how to successfully incorporate spatial information and large receptive field. A recently published paper, Bilateral Segmentation Network for Real-time Semantic Segmentation (BiSeNet) is one of them with relatively good performance and speed. In this semester project, we investigated this paper and derived solutions with multi-scale information incorporation.