Hybrid Stereo Dense Depth Estimation for Robotic Tasks in Industrial Automation

Abstract

We introduce a simple yet effective approach for dense depth reconstruction that operates directly on raw disparity data, eliminating the need for additional disparity refinement stages. By leveraging disparity maps generated from conventional stereo methods, we train a U-Net-based model to directly map disparity to depth, bypassing complex feature engineering. Our method capitalizes on the robustness of traditional stereo matching techniques to varying scenes, focusing exclusively on dense depth reconstruction. This approach not only simplifies the training process but also significantly reduces the requirement for large-scale training datasets. Extensive evaluations demonstrate that our method surpasses classical stereo matching frameworks and state-of-the-art classical post-refinement techniques, achieving superior accuracy. Additionally, our approach offers competitive inference times, comparable to classical as well as end-to-end deep learning methods, making it highly suitable for real-time robotic applications.