Article directory
In recent years, the most widely used large-scale datasets in MVS papers are DTU dataset, Tanks and Temples dataset, ETH3D dataset, and dataset.
For learning-based MVS training, depth maps are essential, while evaluation is based on point clouds. In deep learning for planar scan-based multi-view stereo vision techniques, if a dataset does not contain ground-truth camera calibration, or uses open-source software to obtain ground-truth calibration, then it may not be suitable for training, because planar scan has a significant impact on camera calibration. noise is very sensitive.
1 DTU data set
The DTU dataset is a large-scale MVS dataset published by Aanæs et al. 2106, collected in a controlled laboratory environment with precise camera trajectories.
It contains 124 scenes with 49 or 64 views under 7 different lighting conditions, a total of 128 sets of pictures, each set of data consists of RGB images and corresponding camera parameters, and provides ground truth point clouds scanned using structured light . This dataset is the most popular training and evaluation set in MVS papers in recent years.