Scene Reconstruction Beyond Structure-from-Motion and Multi-View Stereo

Image-based 3D reconstruction has become a robust technology for recovering accurate and realistic models of real-world objects and scenes. A common pipeline for 3D reconstruction is to first apply Structure-from-Motion (SfM), which recovers relative poses for the input images and sparse geometry for the scene, and then apply Multi-view Stereo (MVS), which estimates a dense depthmap for each image. While this two-stage process is quite effective in many 3D modeling scenarios, there are limits to what can be reconstructed. This dissertation focuses on three particular scenarios where the SfM+MVS pipeline fails and introduces new approaches to accomplish each reconstruction task. First, I introduce a novel method to recover dense surface reconstructions of endoscopic video. In this setting, SfM can generally provide sparse surface structure, but the lack of surface texture as well as complex, changing illumination often causes MVS to fail. To overcome these difficulties, I introduce a method that utilizes SfM both to guide surface reflectance estimation and to regularize shading-based depth reconstruction. I also introduce models of reflectance and illumination that improve the final result. Second, I introduce an approach for augmenting 3D reconstructions from large-scale Internet photo-collections by recovering the 3D position of transient objects --- specifically, people --- in the input imagery. Since no two images can be assumed to capture the same person in the same location, the typical triangulation constraints enjoyed by SfM and MVS cannot be directly applied. I introduce an alternative method to approximately triangulate people who stood in similar locations, aided by a height distribution prior and visibility constraints provided by SfM. The scale of the scene, gravity direction, and per-person ground-surface normals are also recovered. Finally, I introduce the concept of using crowd-sourced imagery to create living 3D reconstructions --- visualizations of real places that include dynamic representations of transient objects. A key difficulty here is that SfM+MVS pipelines often poorly reconstruct ground surfaces given Internet images. To address this, I introduce a volumetric reconstruction approach that leverages scene scale and person placements. Crowd simulation is then employed to add virtual pedestrians to the space and bring the reconstruction "to life."Doctor of Philosoph