Joint temporal context exploitation and active learning for video segmentation

摘要

The segmentation of video, or separating out objects in the foreground, is an important application of pattern recognition and computer vision. Segmentation errors in pattern recognition approaches mainly come from difficulties in selecting maximally informative frames for learning. In this paper, we develop an approach to video segmentation that relies on temporal features by modeling the uncertainty of the distribution of different feature mask forms. We use those uncertainty values for unsupervised active learning. We evaluate our approach on the DAVIS16 annotated video data set and Shining3D dental video data set, and the results show our approach to be more accurate than other video segmentation approaches. (C) 2019 Elsevier Ltd. All rights reserved.