CSRF: A robust 3D reconstruction framework for Crowdsourced Street View Videos

This repository presents the official implementation of CSRF, a novel framework designed to enhance depth prediction robustness in challenging real-world scenarios. Our work introduces two key innovations: a custom Ordinal Depth Loss function and a sophisticated multi-modal data filtering pipeline that leverages predictions from both DSTNet and Segment Anything (SAM). The combination of these approaches significantly improves the accuracy and resilience of depth estimation models by focusing on consistent relative depth relationships and ensuring high-fidelity training data.