Scene Segmentation with Deep Neural Networks
This talk is going to present two state-of-the-art networks for scene segmentation tasks. As effective contextual modelling is essential to disambiguate the image nuances, more effective and efficient networks (DAG-RNN, multi-branched CNN) are proposed to perform context aggregation. They are further integrated with pre-trained CNN to yield the backbone of segmentation networks. Besides, class occurrence frequency distribution is highly imbalanced in scene images. Thus, a novel class-weighted loss is introduced to distribute more attention to rare classes during network learning. Overall, these segmentation networks achieve state-of-the-art performance on public scene parsing benchmarks, and their different behaviors (accuracy and efficiency) shed some light on the application prospects of these segmentation networks.
I'm currently a 4-th year PhD student in Nanyang Technological University, and I'm expected to graduate in the upcoming summer. I've been doing research on computer vision problems (semantic segmentation, pose estimation) throughout my PhD research life. I'm always passionate to design novel, fast as well as accurate (segmentation) algorithms so that they may be embedded into industrial products one day. I, together with my colleagues, have participated in ImageNet Scene Parsing 2016 Challenge, and our preliminary model won the sixth place (out of 23 teams) worldwide.