Structure from Motion (SfM) denotes one of the central problems in computer vision. It deals with the reconstruction of a static scene from an image sequence of a single moving camera. This task is typically divided into two alternating stages: tracking, which tries to identify the camera’s position and orientation with respect to a global coordinate system, and mapping, which uses this information to create a depth map from the current camera frame. There are already numerous approaches in the literature concerning local reconstruction techniques which attempt to create sparse point clouds from selected image features. However, the resulting scene information is often insufficient for many fields of application like robotics or medicine. Therefore, dense reconstruction has become more and more prominent in recent research. In 2011, Newcombe et al. [NLD11] presented a new technique called DTAM (Dense Tracking and Mapping), which was one of the first to create fully dense depth maps based on variational methods. Since then, most of the follow-up work concentrated on performance rather than on qualitative optimization due to DTAM’s limited real-time capability compared to sparse methods. It is therefore the objective of this thesis to improve the quality and robustness of the original DTAM algorithm and extend it to a generalized and modular mathematical framework. In particular, the influence of different constancy assumptions and regularizers will be evaluated and tested under various conditions using multiple benchmark data sets.