Sorting unorganized photo sets for urban reconstruction

Guowei Wan, Noah Snavely, Daniel Cohen-Or, Qian Zheng, Baoquan Chen, Sikun Li

Graphical Models 2012

Structure-Preserving Retargeting of Irregular 3D Architecture
Abstract:
In spite of advanced acquisition technology, consumer cameras remain an attractive means for capturing 3D data. For reconstructing buildings it is easy to obtain large numbers of photos representing complete, all-around coverage of a building; however, such large photos collections are often unordered and unorganized, with unknown viewpoints. We present a method for reconstructing piecewise planar building models based on a near-linear time process that sorts such unorganized collections, quickly creating an image graph, an initial pose for each camera, and a piecewise-planar facade model. Our sorting technique first estimates single-view, piecewise planar geometry from each photo, then merges these single-view models together in an analysis phase that reasons about the global scene geometry. A key contribution of our technique is to perform this reasoning based on a number of typical constraints of buildings. This sorting process results in a piecewise planar model of the scene, a set of good initial camera poses, and a correspondence between photos. This information is useful in itself as an approximate scene model, but also represents a good initialization for structure from motion and multi-view stereo techniques from which refined models can be derived, at greatly reduced computational cost compared to prior techniques.
Paper: Link
Results:

Fig. 3. From left to right: local geometry is estimated from each view, including vanishing points, segmented facades, and a local facade model; segmented facades are rectified and clustered, defining a facade graph which is analyzed to produce a simplified set of clusters; given this simplified facade graph, we generate a piece-wise planar facade model, compute a viewpoint for each image, and derive an image connectivity graph; this approximate model forms a good initialization to SfM and subsequent multi-view stereo methods.


Fig. 5. Several single-view facade reconstructions. The first row shows a set of input images, and the second row shows a partial facade model computed from each image. Red arrows represent the estimated position and orientation of the camera.


Fig. 11. Facade model results for our seven real-world datasets. Each dataset shows selected input images, initial cameras, and the output facade model.


Fig. 14. Comparison of sparse SfM reconstruction results of baseline SfM method (Rome in a Day, top row) and our approach (bottom row).


Thanks: This work was supported in part by 973 Program (2009CB723803), NSFC (60902104, 61025012, 61003190, 61170157), 863 Program (2011AA010500), CAS One Hundred Scholar Program, CAS Visiting Professorship for Senior Int’l Scientists, Shenzhen Science and Technology Foundation (JC201005270329A).
Bibtex:

@article{Wan_gm12,
    author = {Guowei Wan, Noah Snavely, Daniel Cohen-Or, Qian Zheng, Baoquan Chen, Sikun Li},
     title = {Sorting unorganized photo sets for urban reconstruction},
     year = {2012},
     month = { January},
     journal = {Graphical Models},
     volume = {74},
     issue = {1},
     pages = {14-28}
}