Self-Supervised Monocular Depth Estimation Scale Recovery using RANSAC Outlier Removal

Recently, self-supervised method has become an increasingly significant branch of depth estimation task, especially in the realm of autonomous driving applications. However, current per-pixel depth maps yielded from RGB images still suffer from uncertain scale factor generated by the nature of monocular image sequences, which further leads to the insufficiency in practical use. In this work, we first analyze such scale uncertainty both theoretically and practically. Then we perform scale recovery utilizing geometric constraint to estimate accurate scale factor, RANSAC(Random sample consensus) outlier removal is introduced into pipeline to obtain accurate ground point extraction. Adequate experiments on KITTI dataset(dataset generated by an autonomous driving platform built up by KIT and TRINA comprising stereo and optical flow image pairs as well as laser data, distributed to train set and test set on account of deep learning), show that, using only camera height prior, our proposed method, though not relying on additional sensors, is able to achieve accurate scale recovery and outperform existing scale recovery methods.