IIIT HYDERABAD Image-based walkthroughs from partial and incremental scene reconstructions Kumar Srijan Syed Ahsan Ishtiaque C. V. Jawahar Center for Visual Information Technology, IIIT-Hyderabad http://cvit.iiit.ac.in Sudipta N. Sinha Microsoft Research, Redmond http://research.microsoft.com Introduction IIIT HYDERABAD Problem IIIT HYDERABAD • Efficiently organize and browse these huge image collections? • Keep Incorporating an incoming stream of images into an existing framework? Related Work IIIT HYDERABAD • • • • • • • • World-Wide Media Exchange (WWMX) PhotoCompas Realityflythrough Aspen Movie Map Photowalker Sea of Images Google Streetview Photo Tourism Photo Tourism IIIT HYDERABAD Input Images Computing correspondences Detect Match keypoints Features in between each each image pair of images Incremental SfM Select a good initial pair to seed reconstruction For each pair, estimate an F-matrix and refine matches Add new images and triangulate new points Full Scene Reconstruction Snavely et. al, Photo Tourism: Exploring image collections in 3D Chain pairwise matches into tracks Bundle adjust Bottlenecks and Issues IIIT HYDERABAD • Global scene reconstruction via incremental structure from motion (Sfm) – Sensitivity to the choice of the initial pair – Cascading of errors – O(N4) in the worst case Snavely et. al, Photo Tourism: Exploring image collections in 3D Bottlenecks and Issues IIIT HYDERABAD • Timing Breakdown Full Scene Reconstruction for Trafalgar Square dataset with 8000 images took > 50 days Snavely et. al, Photo Tourism: Exploring image collections in 3D Our approach IIIT HYDERABAD • “ In a walkthrough, users primarily observe near by overlapping images.” Independent Partial Scene Reconstructions instead of Global Scene Reconstruction • Advantages: – – – – Robustness to errors in incremental SfM module Worst case linear running time Scalable Incremental Partial Reconstructions IIIT HYDERABAD Compute Matches Match Image Refine Matches Correct Match Incorrect Match Standard SfM Compute partial Reconstructions User interface and navigation IIIT HYDERABAD Sample image Input images Verified neighbors Visualization Interface Partial reconstruction Global vs. Partial IIIT HYDERABAD • Global : Allows transition to any image • Partial : Allows transition to a limited number of overlapping images • A -> B implies B -> A A A B B Incremental insertion IIIT HYDERABAD Geometric Match Verification Compute Partial Scene Reconstruction New Image Improve Connectivity Dataset IIIT HYDERABAD Golconda Fort, Hyderabad Fort Dataset 5989 images Results IIIT HYDERABAD Results IIIT HYDERABAD Results IIIT HYDERABAD • Courtyard Dataset with 687 images • Initialized with 200 images • Added 487 image one by one • Largest CC of 674 images. Conclusion IIIT HYDERABAD • Image navigation system based on partial reconstructions can effectively be used to navigate through large collections of images. • Robustness to errors • Able incorporate more images as they become available. Future Work IIIT HYDERABAD • Complete automation – Download images directly from the internet – Add into the framework Acknowledgements IIIT HYDERABAD • “Photo tourism: Exploring photo collections in 3D“ – Noah Snavely, Cornell University – Steven M. Seitz, University of Washington – Richard Szeliski, Microsoft Research Acknowledgements IIIT HYDERABAD • “Visual Word based Location Recognition in 3D models using Distance Augmented Weighting” – Friedrich Fraundorfer, Marc Pollefeys ETH Zürich – Changchang Wu ,JanMichael Frahm ,Marc Pollefeys - UNC Chapel Hill Thank You IIIT HYDERABAD • Questions