An Evaluation of Stereo and Multiview Algorithms for 3d Reconstruction with Synthetic Data

Abstract

The reconstruction of 3D scenes from images has usually been addressed with two different strategies, namely stereo and multi-view. The former requires rectified images and generates a disparity map, while the latter relies on the camera parameters and directly retrieves a depth map. For both cases, deep learning architectures have shown an outstanding performance. However, due to the differences between input and output data, the two strategies are difficult to compare on a common scene. Moreover, for remote sensing applications multi-view data is hard to acquire and the ground truth is either sparse or affected by outliers. Hence, in this article we evaluate the performance of stereo and multi-view architectures trained on synthetic data resembling remote sensing images. The data has been and processed and organized to be compatible with both kind of neural networks. For a fair comparison, training and testing are done only with two views. We focus on the accuracy of the reconstruction, as well as the impact of the depth range and the baseline of the stereo array. Results are presented for deep learning architectures and non-learning algorithms.