Underwater Multiview Stereo Using Axial Camera Models

3D models, generated from underwater imagery, are a valuable asset for many applications. When acquiring images underwater, light is refracted as it passes the boundary layers between water, housing and the air inside the housing due to the different refractive indices of the materials. Thus the geometry of the light rays changes in this scenario and the standard pinhole camera model is not applicable. As a result, pinhole 3D reconstruction methods can not easily be applied in this environment. For the dense reconstruction of scene surfaces the added complexity is especially challenging, as these types of algorithms have to match vast amounts of image content. This work proposes the refractive adaptation of a PatchMatch Multi-View Stereo algorithm. The refraction encountered at flat port underwater housings is explicitly modeled to avoid systematic errors in the reconstruction. Concepts derived from the axial camera model are employed to handle the high demands of Multi-View Stereo regarding accuracy and computational complexity. Numerical simulations and reconstruction results on synthetically generated but realistic images with ground truth validate the effectiveness of the approach.