RANSAC-Scaled Depth: A Dual-Teacher Framework for Metric Depth Annotation in Data-Scarce Scenarios

This paper addresses the problem of training metric monocular depth estimation models for specialized domains in the absence of labeled real-world data. We propose a hybrid pseudo-labeling method that combines the predictions of two models: a metric "teacher," trained on synthetic data to obtain the correct scale, and a foundational relative "teacher" for structurally accurate scene geometry and depth. The relative depth map is calibrated via a linear transformation, whose parameters are found using the outlier-robust RANSAC algorithm on a subset of "support" points. Experiments on the KITTI dataset show that the proposed approach improves the quality of the pseudo-labels, reducing the commonly used error metric AbsRel by 21.6 % compared to the baseline method. A compact "student" model trained on these labels demonstrated superiority over the baseline model, achieving a 23.8 % reduction in AbsRel and a 13.8 % reduction in RMSE log. The results confirm that the proposed method significantly improves domain adaptation from general purpose to the specific domain, allowing for the creation of high-precision metric models without the need to collect and annotate volumes of real data.