Terrain Synthesis by Digital Elevation Models

1. Terrain Synthesis by Digital Elevation Models Howard Zhou, Jie Sun, Greg Turk, and James M. Rehg 2006.10.05

2. Introduction Feature extraction Feature matching Patch stitching Conclusion Table of Contents • Introduction • Feature extraction • Feature matching and alignment • Patch stitching • Conclusion

3. Introduction Feature extraction Feature matching Patch stitching Conclusion Why? • Numerous applications • Landscape design • Flight simulators • Feature film special effects • Computer games

4. Introduction Feature extraction Feature matching Patch stitching Conclusion Previous Work • Terrain synthesis • Fractal model • fBm - fractional Brownian motion (Mandelbrot 1982) • Midpoint displacement, recursive subdivision …(Fournier 1982, Miller 1986, Voss 1985, Lewis 1987, Szeliski, et al. 1989) • Erosion model • Physical erosion simulation (Kelley, et al. 1988) • Combination of Both • Fractal terrains with erosion features (Musgrave et al. 1989) • Most commercial landscaping software such as Terragen, Bryce, Vue d’seprit, and Mojoworld, etc.

5. Introduction Feature extraction Feature matching Patch stitching Conclusion Previous Work • Limitation of previous terrain synthesis approaches • Limited control by user (parameter tuning) • Hard to capture real terrain style

6. Introduction Feature extraction Feature matching Patch stitching Conclusion What If ?

8. Introduction Feature extraction Feature matching Patch stitching Conclusion Related Work • Patch based texture synthesis • Image quilting (Efros and Freeman 2001), Graphcut (Kwatra et al. 2003) • Feature guided texture synthesis • Image analogy (Hertzmann et al. 2000), Feature matching and deformation (Zhang et al. 2003, Wu and Yu 2004)

9. Introduction Feature extraction Feature matching Patch stitching Conclusion Terrain synthesis is not texture synthesis • Terrain synthesis is not simply texture synthesis on height fields. • Terrain synthesis must preserve global features such as ridges and valleys. • Terrain synthesis must be globally controllable. • Unlike general textures, terrain doesn’t have natural boundaries.

10. Introduction Feature extraction Feature matching Patch stitching Conclusion Our Contribution • First example-based terrain synthesis • User control via feature sketches. • Feature-based approach to matching and placement of large curvilinear terrain features. • Tree-ordered patch placement algorithm. • Multiple terrain style.

11. Introduction Feature extraction Feature matching Patch stitching Conclusion Procedure • Feature Extraction • Extract important terrain features (valleys, ridges, …) • Feature matching and deformation • Match terrain features between user sketch and terrain data to find candidate patch • Use deformation to align features • Patch stitching • Use graph cuts and Poisson interpolation to remove visible seams between neighboring patches

12. Introduction Feature extraction Feature matching Patch stitching Conclusion Flowchart Patch stitching Feature extraction Matching and deformation

13. Introduction Feature extraction Feature matching Patch stitching Conclusion Feature Extraction • Finding ridges and valleys • Branches and Ends • Path Features • Chang’s PPA algorithm (Profile recognition and polygon breaking) End Branch Path

14. Introduction Feature extraction Feature matching Patch stitching Conclusion Why PPA? Grand Canyon (shaded relief) Edge detection result PPA result

15. Introduction Feature extraction Feature matching Patch stitching Conclusion PPA explained Profile Recognition Target Connection Polygon Breaking Branch Reduction

16. In action Input

17. In action Profile Recognition

18. In action Polygon building

19. In action Polygon Breaking

20. In action Branch Reduction

21. In action Result

22. Introduction Feature extraction Feature matching Patch stitching Conclusion Feature placement (tree traversal)

23. Introduction Feature extraction Feature matching Patch stitching Conclusion Why is order important? Raster-scan patch placement (ncc) Tree traversal

24. Introduction Feature extraction Feature matching Patch stitching Conclusion Feature alignment • Most of the time, the feature patches need alignment before they can be used. • Thin plate spline mapping for feature deformation • Two sets of corresponding feature points from feature matching • Small deformation in terrain does not alter terrain style Branch End Path

25. Introduction Feature extraction Feature matching Patch stitching Conclusion Feature Patch Matching • d : Deformation energy from TPS warping • g : Graphcut seam cost • f : Feature dissimilarity • i : Other user specified constraints

26. Introduction Feature extraction Feature matching Patch stitching Conclusion Non-feature placement • SSD- based search(accelerated) • Fill the synthesized height map

27. Introduction Feature extraction Feature matching Patch stitching Conclusion Graphcut • Graphcut Textures:Image and Video Synthesis Using Graph Cuts (Kwatra et al. 2003)

28. Introduction Feature extraction Feature matching Patch stitching Conclusion Poisson interpolation • Poisson image editing (Perez et al. 2003) • Modify the gradient and reconstruct

29. Mount Jackson

31. Grand Canyon

33. Flathead National Forest, MT

35. Mount Vernon, KY

38. Middle Earth

44. Introduction Feature extraction Feature matching Patch stitching Conclusion Conclusion • We’ve presented an image-based algorithm for terrain synthesis • It provides user control by intuitive sketch • It preserves terrain style embedded in the original height field

45. Cape Girardeau, MO (failed)

47. Results • Show video

48. Introduction Feature extraction Feature matching Patch stitching Conclusion PPA in action