Accurate 3D Model Generation from Multiple Images

To automatically compute accurate three-dimensional (3D) models from sequences of two-dimensional (2D) handheld images or video, Stellar Science proposes to develop, demonstrate, and evaluate the MOCHI (Models Computed from Handheld Imagery) system. While MOCHI will leverage open source and prior structure from motion (SFM) components developed for the Navy and Air Force to estimate an initial model, the key innovation of MOCHI is the integration of automated and user-assisted methods for computing models robustly under a variety of imaging conditions. This will enable MOCHI to model smooth surfaces where traditional feature-based methods fail and improve the robustness of the recovered geometry in the presence of feature tracking, matching, or voxel carving errors. We will develop standardized metrics for evaluating the accuracy and completeness of the computed model to evaluate MOCHI’s performance. During Phase II we will develop methods to automate 3D model creation from imagery from the proposed test scenarios, segment images using machine learning techniques more accurately, incorporate the ability to interactively sketch objects in 3D to model unseen parts of the object, and determine accuracy of computed models with respect to ground truth.