Reconstruction and visualization of large photo collections

Опубликовано 6 сентября 2016, 17:05

In the Photo Tourism project, we demonstrated the potential for leveraging large, unstructured Internet photo collections to produce compelling 3D visualizations of real-world scenes, through a combination of new computer vision and 3D interaction techniques. More recently, we have extended these techniques in several directions. In this talk, I will describe a new 3D photo browser which addresses some of the challenges in providing effective navigation controls for 3D scenes represented as a set of photos. Our new system gives a user more continuous controls than Photo Tourism, but also analyzes the distribution of photos in a collection to determine a good set of controls for exploring a particular scene. I will also talk about recent work on scaling up 3D reconstruction techniques to handle very large photo collections. Our new reconstruction algorithm improves efficiency by an order of magnitude compared to previous algorithms, and enables reconstruction of large-scale scenes containing several thousand photos.