%0 Conference Paper %1 2019-godard %A Godard, Clement %A Aodha, Oisin Mac %A Firman, Michael %A Brostow, Gabriel %B 2019 IEEE/CVF International Conference on Computer Vision (ICCV) %D 2019 %K depth monocular monodepth2 niantic self-supervised %P 3827-3837 %R 10.1109/ICCV.2019.00393 %T Digging Into Self-Supervised Monocular Depth Estimation %U https://ieeexplore.ieee.org/document/9009796/ %X Per-pixel ground-truth depth data is challenging to acquire at scale. To overcome this limitation, self-supervised learning has emerged as a promising alternative for training models to perform monocular depth estimation. In this paper, we propose a set of improvements, which together result in both quantitatively and qualitatively improved depth maps compared to competing self-supervised methods. Research on self-supervised monocular training usually explores increasingly complex architectures, loss functions, and image formation models, all of which have recently helped to close the gap with fully-supervised methods. We show that a surprisingly simple model, and associated design choices, lead to superior predictions. In particular, we propose (i) a minimum reprojection loss, designed to robustly handle occlusions, (ii) a full-resolution multi-scale sampling method that reduces visual artifacts, and (iii) an auto-masking loss to ignore training pixels that violate camera motion assumptions. We demonstrate the effectiveness of each component in isolation, and show high quality, state-of-the-art results on the KITTI benchmark.

@inproceedings{2019-godard, abstract = {Per-pixel ground-truth depth data is challenging to acquire at scale. To overcome this limitation, self-supervised learning has emerged as a promising alternative for training models to perform monocular depth estimation. In this paper, we propose a set of improvements, which together result in both quantitatively and qualitatively improved depth maps compared to competing self-supervised methods. Research on self-supervised monocular training usually explores increasingly complex architectures, loss functions, and image formation models, all of which have recently helped to close the gap with fully-supervised methods. We show that a surprisingly simple model, and associated design choices, lead to superior predictions. In particular, we propose (i) a minimum reprojection loss, designed to robustly handle occlusions, (ii) a full-resolution multi-scale sampling method that reduces visual artifacts, and (iii) an auto-masking loss to ignore training pixels that violate camera motion assumptions. We demonstrate the effectiveness of each component in isolation, and show high quality, state-of-the-art results on the KITTI benchmark.}, added-at = {2021-07-05T22:33:18.000+0200}, author = {Godard, Clement and Aodha, Oisin Mac and Firman, Michael and Brostow, Gabriel}, biburl = {https://www.bibsonomy.org/bibtex/24bde975f5bcce1b4a04282fa59e2a4d4/pkoch}, booktitle = {2019 IEEE/CVF International Conference on Computer Vision (ICCV)}, doi = {10.1109/ICCV.2019.00393}, interhash = {6b006f54c34fc4aa8bfe493e0db6212a}, intrahash = {4bde975f5bcce1b4a04282fa59e2a4d4}, issn = {2380-7504}, keywords = {depth monocular monodepth2 niantic self-supervised}, month = oct, pages = {3827-3837}, timestamp = {2021-07-05T22:33:18.000+0200}, title = {Digging Into Self-Supervised Monocular Depth Estimation}, url = {https://ieeexplore.ieee.org/document/9009796/}, year = 2019 }