Dense 3D modeling based on monocular visual data is a powerful process of gaining spatial 3D understanding from 2D observations. The use of visual data to reconstruct such 3D models is still a challenging topic. To obtain the accurate dimensions, additional metadata is required such as a GPS which is not always available. Besides this, dealing with challenging visual situations such as bad-lightening conditions or motion blur remains a difficult subject. Furthermore, since visual data is highly dimensional, most algorithms lack scalability, meaning that they fail to reconstruct 3D models in acceptable time limits and are incapable of handling large data sets. In this thesis, the objective is to mitigate these typical issues whilst preserving the quality of the dense 3D models. To this end, visual-inertial SLAM and MVS techniques are combined to form a dense 3D modeling architecture that is capable of mitigating the typical challenges of classical dense 3D modeling approaches. Besides this, the architecture is extended with additional improvements in the MVS system. These improvements further increase the scalability by abstracting the input data in a highly compact representation by leveraging image segmentation techniques. The result is a novel, visual-inertial dense 3D modeling system. The novel system is tested on benchmark data sets and within a lab setting, where a remote-inspection case-study is performed. The presented system is compared against the industrial and academic state-of-the-art systems. A thorough comparison is made by evaluating the pose accuracy, computation time, and reconstruction quality. It is shown that the presented system improves the state-of-the art systems by a significant margin in terms of computation-time. Furthermore, the presented system is capable of computing 3D models with accurate geometric scale without relying on external metadata, showcasing the effectiveness of the presented system. This work contributes to the lack of research in dense 3D modeling based on visual-inertial SLAM and paves the way for a new direction of efficient MVS algorithms.