Scene Reconstruction from Community Photo Collections

The Internet is an unprecedented source of visual information about our world, and has recently sparked a great deal of interest in computer vision and computer graphics. This short course describes how we can extract geometry and appearance from massive, unstructured Internet collections, through new advances in structure from motion, multi-view stereo, and appearance acquisition.

Course Slides as PDF

• Introduction (Goesele)
• The Photo Tourism system for 3D scene reconstruction and Using Photo Tourism to visualize and explore scenes (Snavely)
• Multi-View Stereo (Goesele)
• Appearance Acquisition (Lensch)
• Conclusion (all)

Course Content

Introduction

• Course overview
• Data sources and tools for community photo collections (CPC)

The Photo Tourism system for 3D scene reconstruction (Snavely)

• Basic reconstruction pipeline
  • Finding the images
  • Feature detection and matching
  • Structure from motion
  • Essential algorithms for reconstruction
  • Techniques for robustness and stability
• Optimization techniques for large-scale bundle adjustment
  • Sparse direct solvers
  • Conjugate gradient algorithms

Using Photo Tourism to visualize and explore scenes (Snavely)

• Morphing and geometric proxies
• Navigation controls
• Photosynth

Multi-View Stereo (Goesele)

• Geometry reconstruction using stereo and multi-view stereo techniques
• Reconstruction challenges
• Multi-view stereo for CPCs
• Geometry post-processing

Appearance Acquisition (Lensch)

• Introduction to appearance representation and rendering
• Illumination estimation
• Reflectance estimation
• Implicit appearance representation

Conclusion

External Resources

• Community Photo Collections at the University of Washington
• Photo Tourism