Probabilistic Scene Modeling from Aerial Imagery

Daniel Crispell, Joseph Mundy, and Gabriel Taubin

Brown University

The goal of this project is to provide a means for the automatic reconstruction of probabilistic 3-d models from aerial imagery which fully represent the uncertainties and ambiguities inherent in the available data. The models are capable of providing rich information about the represented scene, including visibility and occlusion information for points in the scene, with a focus on the generation of "expected images" rendered from virtual viewpoints.

Our method represents uncertain geometry using a continuous density field. Visibility probabilities and expected intensity values can be computed by integrating along rays through the volume. The density is sampled using an octree, allowing efficient representation of large and complex outdoor scenes not possible with other probabilistic methods.


results: downtown sequence
downtown sfm results
downtown uncertain geometry
downtown virtual views
Input camera positions (calibrated automatically using a Structure from Motion algorithm)
The reconstructed probabilistic geometry (visualized using volume rendering software)
Positions of the cameras from which virtual views are generated


interactive display of virtual views (click and drag to move camera, may take a minute to download)




results: capitol sequence
downtown sfm results
downtown uncertain geometry
downtown virtual views
Input camera positions (calibrated automatically using a Structure from Motion algorithm)
The reconstructed probabilistic geometry (visualized using volume rendering software)
Positions of the cameras from which virtual views are generated


Interactive display of virtual views (click and drag to move camera, may take a minute to download)