Deep monocular depth estimation leveraging a large-scale outdoor stereo dataset

作者：

Highlights：

• A novel framework for monocular depth estimation via a student–teacher strategy.

• Introducing a data ensemble and stereo confidence-guided regression loss.

• Constructing a new large-scale outdoor stereo dataset named the DIML/CVL dataset.

• Demonstrating the feature representation of our trained-model for high-level tasks.

摘要

•A novel framework for monocular depth estimation via a student–teacher strategy.•Introducing a data ensemble and stereo confidence-guided regression loss.•Constructing a new large-scale outdoor stereo dataset named the DIML/CVL dataset.•Demonstrating the feature representation of our trained-model for high-level tasks.

论文关键词：Monocular depth estimation,Convolutional neural network,Student–teacher strategy,Outdoor stereo dataset,Stereo confidence maps

论文评审过程：Received 21 September 2020, Revised 2 January 2021, Accepted 5 March 2021, Available online 16 March 2021, Version of Record 23 April 2021.

论文官网地址：https://doi.org/10.1016/j.eswa.2021.114877