Depth Super Resolution by Rigid Body Self-Similarity in 3D

摘要: We tackle the problem of jointly increasing spatial resolution and apparent measurement accuracy an input low-resolution, noisy, perhaps heavily quantized depth map. In stark contrast to earlier work, we make no use ancillary data like a color image at target resolution, multiple aligned maps, or database high-resolution exemplars. Instead, proceed by identifying merging patch correspondences within map itself, exploiting wise scene self-similarity across such as repetition geometric primitives object symmetry. While notion 'single-image' super has successfully been applied in context intensity images, are our knowledge first present tailored analogue for images. Rather than reason terms patches 2D pixels others have before us, key contribution is reasoning 3D points, with matched pairs related respective 6 DoF rigid body motion 3D. support obtaining dense correspondence field reasonable time, introduce new variant Patch Match. A third simple, yet effective up scaling technique, which predicts sharp boundaries resolution. show that results highly competitive those alternative techniques leveraging even