Monocular Depth Estimation Through Virtual-World Supervision and Real-World SfM Self-Supervision

Depth information is essential for on-board perception in autonomous driving and driver assistance. Monocular depth estimation (MDE) very appealing since it allows appearance being on direct pixelwise correspondence without further calibration. Best MDE models are based Convolutional Neural Networks (CNNs) trained a supervised manner, i.e ., assuming ground truth (GT). Usually, this GT acquired at training time through calibrated multi-modal suite of sensors. However, also using only monocular system cheaper more scalable. This possible by relying structure-from-motion (SfM) principles to generate self-supervision. Nevertheless, problems camouflaged objects, visibility changes, static-camera intervals, textureless areas, scale ambiguity, diminish the usefulness such In paper, we perform xmlns:xlink="http://www.w3.org/1999/xlink">mono cular ${d}$ epth notation="LaTeX">${e}$ stimation notation="LaTeX">${v}$ irtual-world notation="LaTeX">${s}$ upervision (MonoDEVS) real-world SfM We compensate self-supervision limitations leveraging virtual-world images with accurate semantic supervision, addressing virtual-to-real domain gap. Our MonoDEVSNet outperforms previous CNNs even stereo sequences.