de ghosting for gigapixel snapshot processing
play

De-ghosting for Gigapixel Snapshot Processing Alexandros-Stavros - PowerPoint PPT Presentation

De-ghosting for Gigapixel Snapshot Processing Alexandros-Stavros Iliopoulos 1 Jun Hu 1 Nikos Pitsianis 2 , 1 Xiaobai Sun 1 Mike Gehm 3 David Brady 1 1 Duke University 2 Aristotle University of Thessaloniki 3 University of Arizona March 20, 2013


  1. De-ghosting for Gigapixel Snapshot Processing Alexandros-Stavros Iliopoulos 1 Jun Hu 1 Nikos Pitsianis 2 , 1 Xiaobai Sun 1 Mike Gehm 3 David Brady 1 1 Duke University 2 Aristotle University of Thessaloniki 3 University of Arizona March 20, 2013

  2. Introduction De-ghosting Recap Acknowledgments References Outline 1 Introduction 2 De-ghosting Pipeline Alignment Fusion Illustrations 3 Recap 4 Acknowledgments A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 2/36

  3. Introduction De-ghosting Recap Acknowledgments References Example Multi-Camera Systems Higher-end performance through lower-end cameras System Overlap ratio Purpose Ref. high frame-rate video; 1 ∼ 90% Stanford Multi-Camera Array (mode 1) synthetic aperture 1 Stanford Multi-Camera Array (mode 2) ∼ 50% high resolution eFOV 2 , 3 AWARE-2 ∼ 10% high resolution eFOV 4 ARGUS-IS ∼ 5% high resolution eFOV 5 Single-camera sweep over stationary scene variable high resolution eFOV Overlap large small A B C D 1 B. Wilburn et al . ACM Transactions on Graphics 24:3, 2005. 2 D.J. Brady et al . Nature 486:7403, 2012. 3 F.R. Golish et al . Optics Express 20:20, 2012. 4 B. Leininger et al . SPIE 6981, 2008. 5 J. Kopf et al . ACM Transactions on Graphics 26:3, 2007. A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 3/36

  4. Introduction De-ghosting Recap Acknowledgments References AWARE-2 Prototype: 2 Gigapixels, 120 o FOV Independent focus & exposure Gigapixel-resolution snapshots Complex configuration on a hemisphere D.J. Brady et al . Nature 486:7403, 2012. D.R. Golish et al . Optics Express 20:20, 2012. E.J. Tremblay et al . Applied Optics 51:20, 2012. AWARE-2 image acquisition outline. Image taken from http://www.mosaic.disp.duke.edu/AWARE/index.html . A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 4/36

  5. Introduction De-ghosting Recap Acknowledgments References Gigapixel Imaging Applications Survey, query and monitoring of: urban and suburban development 1 1 M.A. Smith. Fine International Conference on Gi- wild-life habitats 2 gapixel Imaging for Science , 2010. 2 M.H. Nichols et al . Rangeland Ecology & Man- archaeological sites 3 agement 62, 2009. 3 M. Seidl and C. Breiteneder. VAST , 2011. 4 A. McEwen et al . Journal of Geophysical Research: Exploration and dynamics of celestial Planets 115, 2007. 5 L. Gueguen et al . IGARSS , 2011. bodies 4 6 B. Leiningen et al . SPIE 6981, 2008. Recognition 5 Surveillance 6 A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 5/36

  6. Introduction De-ghosting Recap Acknowledgments References Stitching Software GigaPan Stitch 1 Overlap large small Autopano Giga 2 Microsoft ICE 3 Autostitch 4 Configuration geometry Panorama Tools 5 e.g. MS ICE, Autopano e.g. GigaPan Stich (Cartesian grid) Fiji 6 Free-form Pre-mandated ... Challenged by complex, sparse Customized geometry & small, noisy overlap 1 gigapan.com/ 2 autopano.net/ 3 research.microsoft.com/en-us/UM/redmond/groups/IVM/ICE/ 4 www.cs.bath.ac.uk/brown/autostitch/autostitch.html 5 panotools.sourceforge.net/ 6 http://fiji.sc/ A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 6/36

  7. Introduction De-ghosting Recap Acknowledgments References FoV Overlap: Small, Sparse, Noisy Note: AWARE-10 is coming out; see M. Gehm’s talk A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 7/36

  8. Introduction De-ghosting Recap Acknowledgments References FoV Overlap: Small, Sparse, Noisy Note: AWARE-10 is coming out; see M. Gehm’s talk A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 7/36

  9. Introduction De-ghosting Recap Acknowledgments References Outline 1 Introduction 2 De-ghosting Pipeline Alignment Fusion Illustrations 3 Recap 4 Acknowledgments A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 8/36

  10. Introduction De-ghosting Recap Acknowledgments References Outline 1 Introduction 2 De-ghosting Pipeline Alignment Fusion Illustrations 3 Recap 4 Acknowledgments A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 9/36

  11. Introduction De-ghosting Recap Acknowledgments References Ghosting & De-ghosting Ghosted image De-ghosted using our pipeline Both results from the AWARE-2 (monochrome) dataset (AWARE-10 produces color images) A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 10/36

  12. Introduction De-ghosting Recap Acknowledgments References Ghost Sources Static/systematic: Deviations from design during manufacturing Displacement in array mounting Transient/scene-dependent: Variable camera viewpoints Independent camera parameters & settings A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 11/36

  13. Introduction De-ghosting Recap Acknowledgments References De-ghosting: 3 Key Steps (simultaneous transformations) Pairwise registration (control point matching) Global bundle adjustment among multiple images Gradient-domain blending (merged gradients) (blended image) A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 12/36

  14. Introduction De-ghosting Recap Acknowledgments References De-ghosting Pipeline Raw ¡Images, ¡ Geometric ¡ Fusion ¡ Flat-­‑fields ¡ Alignment ¡ Approximate ¡ Reliable ¡ Feature ¡ Global ¡ Gradient ¡ Gradient ¡ Overlapping ¡ Feature ¡ Extrac8on ¡ Bundle ¡ Merging ¡ Integra8on ¡ Regions ¡ Matching ¡ Block ¡Operator ¡ Laplacian ¡Solver ¡ Pixel-­‑wise ¡Operator ¡ A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 13/36

  15. Introduction De-ghosting Recap Acknowledgments References Outline 1 Introduction 2 De-ghosting Pipeline Alignment Fusion Illustrations 3 Recap 4 Acknowledgments A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 14/36

  16. Introduction De-ghosting Recap Acknowledgments References Pairwise Registration Sparse, Small, Noisy overlapping regions Geometric computation-intensive SIFT anchor points SiftGPU by C.C. Wu 1 configuration “broken” ghosted reliable control points GeCo-RANSAC preconditioning Global Bundle Adjustment 1 http://cs.unc.edu/~ccwu/siftgpu A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 15/36

  17. Introduction De-ghosting Recap Acknowledgments References Bundle Adjustment Adhere to geometric configuration 15 15 14 14 ∑︂ ∑︂ ⃦ x T ⃦ k , i H i − x T ⃦ min w ij k , j H j 13 13 (variational form 1) 16 16 ⃦ 6 6 2 { H i } 5 5 I i ∩ I j ̸ = ∅ x k ∈ℳ ij 7 7 2 2 R R R 12 12 min H ‖ WE x H ‖ 2 (variational form 2) 4 4 3 3 8 8 11 11 10 10 Fix a reference frame R : 9 9 L ¯ R H ¯ R = B R (normal/Laplace equation) strong overlap weak overlap A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 16/36

  18. Introduction De-ghosting Recap Acknowledgments References Outline 1 Introduction 2 De-ghosting Pipeline Alignment Fusion Illustrations 3 Recap 4 Acknowledgments A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 17/36

  19. Introduction De-ghosting Recap Acknowledgments References Gradient Re-projection Place & compute gradients on the mosaic canvas Pack images into non-overlapping pairs Custom CUDA kernels Transformation back-projection; interpolation Binary image erosion to remove spurious gradient border Speed-up by packing & GPU: 40x A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 18/36

  20. Introduction De-ghosting Recap Acknowledgments References Gradient-domain Blending Maintains high-frequency information Smooths intensity seams Invariant to camera sensor bias Computation-intensive integration ∑︂ ∇ I ( x ) = w i ( x ) ∇ I i ( x ) x ∈ I i I = G * div( ∇ I ) Green’s function ( G ) is approximated via a convolution pyramid. 1 Speed-up by algorithm, memory streaming, GPU: 30x 1 Z. Farbman et al . ACM Transactions on Graphics 30, 2011. A.S. Iliopoulos, J. Hu, N. Pitsianis, X. Sun, M. Gehm, D. Brady Duke, AUTh, Arizona De-ghosting for Gigapixel Snapshot Processing 19/36

Recommend


More recommend