On the benefit of adversarial training for monocular depth estimation

作者：

Highlights：

• GANs are useful when the reconstruction loss is not too constrained.

• When using a constrained loss, non-adversarial training outperforms GAN training.

• Evaluating reconstruction losses at many scales over-regularises the final disparity.

• Batch normalisation improves depth prediction quality significantly.

摘要

•GANs are useful when the reconstruction loss is not too constrained.•When using a constrained loss, non-adversarial training outperforms GAN training.•Evaluating reconstruction losses at many scales over-regularises the final disparity.•Batch normalisation improves depth prediction quality significantly.

论文关键词：Monocular depth estimation,Adversarial training,GAN

论文评审过程：Received 24 July 2019, Revised 9 October 2019, Accepted 9 October 2019, Available online 18 October 2019, Version of Record 15 November 2019.

论文官网地址：https://doi.org/10.1016/j.cviu.2019.102848