Improving Monocular Depth Prediction with Novel View Synthesis

December 04, 2021

Authors: Zuria Bauer, Zuoyue Li, Sergio Orts-Escolano, Miguel Cazorla, Marc Pollefeys and Martin R Oswald

Conference: 2021 International Conference on 3D Vision (3DV)

Abstract

Building upon the recent progress in novel view synthesis, we propose its application to improve monocular depth estimation. In particular, we propose a novel training method split in three main steps. First, the prediction results of a monocular depth network are warped to an additional view point. Second, we apply an additional image synthesis network, which corrects and improves the quality of the warped RGB image. The output of this network is required to look as similar as possible to the ground-truth view by minimizing the pixel-wise RGB reconstruction error. Third, we reapply the same monocular depth estimation onto the synthesized second view point and ensure that the depth predictions are consistent with the associated ground truth depth. Experimental results prove that our method achieves state-of-the-art or comparable performance on the KITTI and NYU-Depth-v2 datasets with a lightweight and simple vanilla U-Net architecture.

Contributions

• We propose to use novel-view synthesis as an additional supervisory signal to improve the training of a monocular depth estimation network. To this end, we propose two loss functions that augment the traditional depth supervision.
• We present comprehensive experiments on both indoor and outdoor datasets that demonstrate the benefits of our approach, as well as an ablation study which empirically justifies our design choices.

Results

State-of-the-art comparison on the KITTI dataset. For reference, we additionally show the results of our U-Net baseline in the second-to-last row, which is the same network, but trained without the proposed NVS losses. The reported numbers are from the corresponding original papers. Best results are shown in bold and second best results in blue.

State-of-the-art comparison on the NYU-Depth-v2 dataset. Please note the substantial reduction of the relative error by our approach. The reported numbers are from the corresponding original papers. Best results are shown in bold and second best results in blue.

Code