Doctor of Philosophy (PhD)
Division of Computer Science and Engineering
Geometry is the essential property of real-world scenes. Understanding the shape of the object is critical to many computer vision applications. In this dissertation, we explore using computational imaging approaches to recover the geometry of real-world scenes. Computational imaging is an emerging technique that uses the co-designs of image hardware and computational software to expand the capacity of traditional cameras. To tackle face recognition in the uncontrolled environment, we study 2D color image and 3D shape to deal with body movement and self-occlusion. Especially, we use multiple RGB-D cameras to fuse the varying pose and register the front face in a unified coordinate system. The deep color feature and geodesic distance feature have been used to complete face recognition. To handle the underwater image application, we study the angular-spatial encoding and polarization state encoding of light rays using computational imaging devices. Specifically, we use the light field camera to tackle the challenging problem of underwater 3D reconstruction. We leverage the angular sampling of the light field for robust depth estimation. We also develop a fast ray marching algorithm to improve the efficiency of the algorithm. To deal with arbitrary reflectance, we investigate polarimetric imaging and develop polarimetric Helmholtz stereopsis that uses reciprocal polarimetric image pairs for high-fidelity 3D surface reconstruction. We formulate new reciprocity and diffuse/specular polarimetric constraints to recover surface depths and normals using an optimization framework. To recover the 3D shape in the unknown and uncontrolled natural illumination, we use two circularly polarized spotlights to boost the polarization cues corrupted by the environment lighting, as well as to provide photometric cues. To mitigate the effect of uncontrolled environment light in photometric constraints, we estimate a lighting proxy map and iteratively refine the normal and lighting estimation. Through expensive experiments on the simulated and real images, we demonstrate that our proposed computational imaging methods outperform traditional imaging approaches.
Ding, Yuqi, "Computational Imaging for Shape Understanding" (2022). LSU Doctoral Dissertations. 5945.
Karki, Bijaya B.