Unsupervised Monocular Depth Estimation with Left-Right Consistency[CVPR 17]

One Line Summary

Existing approaches treat depth prediction as a supervised regression problem and as a result, require vast quantities of corresponding ground truth depth data for training. This paper replaces the use of explicit depth data during training with easier-to-obtain binocular stereo footage.

Motivation

Just recording quality depth data in a range of environments is a challenging problem. A novel training objective that enables our convolutional neural network to learn to perform single image depth estimation, despite the absence of ground truth depth data

Detailed Summary

• Exploiting epipolar geometry constraints, disparity is generated using the image reconstruction loss.
• Training loss is a weighted sum of apperance matching loss, disparity smoothness loss, left-right disparity consistency loss.
• Exploit the ease with which binocular stereo data can be captured.
• This model can generalize to unseen datasets and still produce visually plausible depth maps

Novelty and Contributions

• L R Consistency

• Training objective that allows the network to estimate single image depth in the absence of ground truth data.

Network Details

• Network uses the hour glass type architecture .
• ResNet 50 is used as the encoder part of the network .
• Decoder part appends the predictions at each resolution for making predictions at the next higher resolution.

Results

Authours

Clément Godard, Oisin Mac Aodha, Gabriel J. Brostow

Sources

Paper

Code