Dense Reconstruction of Underwater Scenes from Monocular Sequences of Images

Dense Euclidean reconstruction of 3D scenes is still a challenging task despite significant progress in the last 10 years. Data acquired underwater by a single, freely moving (handheld) camera is even more difficult to process, because of the lack of reliable salient points, wide and varying baselines between overlapping views, and water blurriness and wavelength-dependent light attenuation. Besides, most of the currently collected underwater imagery has been acquired without any consideration of basic photogrammetric requirements. As neither color nor brightness constancy holds for underwater imagery, the only reliable cue is a texture, which by definition has a spatial extent and changes its spatial frequencies that depend on the direction of view. This paper proposes a novel technique for quasi-dense close-range Euclidean reconstruction that was motivated by ideas developed for photogrammetric applications. The proposed approach starts with a sparse set of highly robust matches (seeds) and expands pair-wise matches into their neighborhood until no more reliable correspondences can be found. The Adaptive Least Square Matching (ALSM) technique is used during the search process to establish new matches in order to increase the robustness of the solution and to avoid mismatches. Experiments on a typical underwater image dataset have demonstrated promising results.