Orientation of images with low contrast textures and transparent objects

Objects with non-collaborative or transparent surfaces pose challenges to image orientation procedures and are an open research task in photogrammetry and computer vision. In this paper, we analyse the critical issues that cause image orientation failures and propose two approaches that leverage the low-contrast textures present on object surfaces to accurately orient an image block. Both approaches privilege tie point detection on low-contrast textures, discarding specular reflections and static tie points. In the first approach local descriptors are extracted in those regions where roughness and micro-structures are better highlighted, applying the normalized cross-correlation (NCC) on the gradient map of the images to fully exploit the geometrical content of the patches. The second approach builds on the first method modifying the classic RootSIFT pipeline and obtaining a faster and more reliable approach. Different transparent objects with different surface characteristics are tested to evaluate the efficiency and reliability of the proposed pipelines for image orientation and successive dense 3D reconstruction.