A complete system of 3D reconstruction from images contains two main components: Structure-from-Motion (SfM) and Multiple View Stereo (MVS). The first component, SfM, recovers camera poses of each image and sparse point positions, and the second component, MVS, recovers 3D representation of scenes or objects in the images. Towards the large scale 3D reconstruction from images, the state of the arts of SfM and MVS suffers from scalability due to memory resources. In this thesis, we propose a systematic approach to accommodate the visual reconstruction scalability for very large scale data-sets.

The large scale optimization is fundamental for Structure-from-Motion. To break through the memory limitation for the large scale global bundle adjustment in SfM, we first propose a dist...[ Read more ]

A complete system of 3D reconstruction from images contains two main components: Structure-from-Motion (SfM) and Multiple View Stereo (MVS). The first component, SfM, recovers camera poses of each image and sparse point positions, and the second component, MVS, recovers 3D representation of scenes or objects in the images. Towards the large scale 3D reconstruction from images, the state of the arts of SfM and MVS suffers from scalability due to memory resources. In this thesis, we propose a systematic approach to accommodate the visual reconstruction scalability for very large scale data-sets.

The large scale optimization is fundamental for Structure-from-Motion. To break through the memory limitation for the large scale global bundle adjustment in SfM, we first propose a distributed method based on space division with the ADMM algorithm, so that the whole SfM can be computed in a distributed manner. Then, we design a new positional measurement fusion method as the application of the large scale optimization to utilize available positional measurements from other sources to improve the accuracy of camera poses.

Multiple View Stereo algorithms require to select and cluster images from large scale redundant image sets, so that they can process the most suitable images under the memory limitation. In this thesis, similar with the proposed distributed optimization framework, we propose a space division based method to select and cluster images to obtain high quality dense point clouds in the dense reconstruction.

All these contributions are critical to the modern 3D reconstruction in very large scale context, and have been demonstrated in large collection of public and private data-sets. By the proposed large scale optimization framework in SfM, and the image selection and clustering method for MVS, we can handle large scale image data-sets in a fully distributed manner in 3D reconstruction to produce accurate 3D representations automatically and efficiently.