Is Real-time Deep Learning-based Monocular Depth Estimation accurate for Multi-Camera Setups?

Depth estimation is a key component for real-time immersive experiences that require volumetric capture systems using RGB+Depth information. Recent developments on deep learning-based monocular depth estimation report promising results, but they are mostly general purpose and they have not been tested for multi-camera volumetric reconstruction. In this work, we present an evaluation of state-of-the-art monocular depth estimation applied to real-time volumetric capture. We propose a framework to evaluate the quality of the generated volumetric information. Selected deep learning based-models will be used to generate depth maps for a multi-camera setup. Their performance will be analyzed using a Free Viewpoint Video (FVV) system which relies on Depth Image Based Rendering (DIBR) to synthesize virtual views of the content. The models will be compared in terms of execution time and quality of the synthesized views.