RealMonoDepth: Self-Supervised Monocular Depth Estimation for General Scenes

• DOI: 10.48550/arxiv.2004.06267
• Publication Date: 2020-01-01
• Authors: Mertalp Ocal; Armin Mustafa
• Keywords: Monocular

We present a generalised self-supervised learning approach for monocular estimation of the real depth across scenes with diverse ranges from 1--100s meters. Existing supervised methods require accurate measurements training. This limitation has led to introduction that are trained on stereo image pairs fixed camera baseline estimate disparity which is transformed given known calibration. Self-supervised approaches have demonstrated impressive results but do not generalise different or baselines. In this paper, we introduce RealMonoDepth learns scene range indoor and outdoor scenes. A novel loss function respect true based relative scaling warping proposed. allows training single network multiple data sets both pair in wild moving sets. comprehensive performance evaluation five benchmark demonstrates provides generalises scenes, consistently outperforming previous approaches.