Information aggregation and fusion in deep neural networks for object interaction exploration for semantic segmentation

摘要

To tackle the semantic segmentation task, which is a fundamental problem in computer vision, various approaches have been proposed. However, how to utilize object interaction information for improving semantic segmentation performances is not paid enough attention to. In this paper, we propose a method for information aggregation and fusion for exploring object interaction information effectively for improving semantic segmentation performances. Specifically, we propose a logit aggregation strategy to explore object interaction information for semantic segmentation. Furthermore, to facilitate object interaction to guide the training of the semantic segmentation model, we propose to fuse features from intermediate layers of the model to aid pixel semantic label predication. And to fuse these features effectively, a buffered layer connection approach is presented. The proposed method is evaluated extensively in experiments. Obtained results demonstrate the effectiveness of the proposed method.