Joint temporal context exploitation and active learning for video segmentation

Highlights：

• In this work, we introduce multitask learning using a recurrent network to improve the accuracy of motion inference in multi-scale analysis.

• We present a novel way to temporal context to infer the mask in the next frame in a probability framework, and compare this result to the spatial result according to the confidence interval.

• We use uncertainty in the probability mask transfer to improve sample selection for unsupervised active learning.

• We also collect and annotate oral video sequence, and create a new data set called Shining3D dental data set, so that the approach proposed in this paper can be compared with other state-of-the-art approaches on video segmentation.