The generation of depth maps is essential for numerous applications, such as autonomous driving or augmented reality. Typically these are generated from stereo image pairs or by making use of active sensors (e.g. LiDAR or RGB-D cameras). Based on the monocular depth perception of humans, this project investigates the estimation of depth maps from single images using artificial neural networks
So far, only visual comparisons and global error metrics have been used for the qualitative assessment of such predicted depth maps. However, they do not consider local geometric properties of the depth maps. For this reason, we present a series of new, geometrically interpretable metrics which concentrate on the following quality criteria of depth maps:
- Accuracy and completeness of depth discontinuities
- Planarity and orientation of planar objects
- Distance-related assessment
Further information on the metrics can be found in this ECCV 2018 WS Paper.
Employing our new error metrics require a high-quality RGB-D reference dataset. With the help of a customized acquisition setup consisting of a DSLR camera and a terrestrial laser scanner, a series of dense and accurately aligned RGB-D image pairs can be generated, which stand out for high resolution and low noise. Our acquired image pairs provide sharp depth discontinuities, as well as noise-free planar surfaces, which are basic requirements for our error metrics.
This setup was used to create our new test data set iBims-1 (independent Benchmark images and matched scans - version 1) for monocular depth estimation. More information about our dataset can be found on this project page.
- Monocular-Depth Assisted Semi-Global-Matching. International Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (ISPRS) (tbp), 2019 more…
- Evaluation of CNN-Based Single-Image Depth Estimation Methods. Proceedings of the European Conference on Computer Vision Workshops (ECCV-WS), Springer International Publishing, 2019 more…