3 D Photography Beyond Perspective Steve Seitz Dept
3 D Photography: Beyond Perspective Steve Seitz Dept. Computer Science & Eng. University of Washington
Perspective Projection Humans evolved with perspective eyes • Capture light along rays that converge at a single point
Cameras also evolved with perspective optics • Optimized for humans, not computers!
A non-perspective image
The Blue Marble, NASA satellite image composite
The Blue Marble, NASA satellite image composite
The Blue Marble, NASA satellite image composite
Cyberware Scanner
Print Gallery, by M. C. Escher, 1956
Panorama from Disney’s 1940 film Pinocchio (from Wood et al. , SIGGRAPH 1997)
What’s an Image? perspective image general image An image is any 2 D subset of rays in space • Actually, the light energy flowing along these rays • Any 2 D “slice” of the plenoptic function
Non-perspective Imaging Issues: • What other types of images are possible? • Which images are useful? • How can we capture these images?
Path Images
Path Images
Path Images y x t
Path Images y x t
Path Images Pushbroom images • satellite Bolles et al. [87] • y EPI Tsuji et al. [92] • omni-directional image Peleg et al. [97] • x t manifold mosaic Radamacher & Bishop [99] • . . . MCOP
Linear Path Video Cube Demo • application by Michael Cohen et al. , Microsoft input pushbroom images
Circular Path circular EPI panorama input image (“concentric mosaic”)
Circular Path input video cube cyclographs
What are these images good for? • Applications to computer graphics, computer vision?
• • Circular Stereo [Ishiguro, Yamamoto, Tsuji, 92] [Peleg and Ben-Ezra, 99] [Shum, Kalai, Seitz, 99] [Nayar and Karmarkar, 00] y x 0 - nx 0 +n x q I
Stereo Panorama Disparity map result
Stereo Panorama Dark--close, light--far
Circular Stereo Advantages • 360 degree scene reconstruction • Uniform accuracy, optimal
Pushbroom Stereo y x 0 - nx 0 +n x t
Stereo Cyclographs y x 0 - nx 0 +n x q
Stereo Cyclograph Reconstruction Computed from two cyclograph images • Using unmodified stereo matcher [Zitnick & Kanade]
Stereo Path Images Do these all produce stereo pairs? • two images with horizontal parallax Yes
Stereo Path Images How about this path? No • Yes if the camera path is a conic – line, circle, ellipse, parabola, hyperbola • Must capture rays lying on doubly-ruled quadrics – [Padja 2001], [Seitz 2001]
Stereo Parabolic Panoramas perspective image parabolic panorama
Beyond Perspective Cameras for humans, not machines! • Need to rethink cameras • Image should suit the task Future: cameras will evolve like CPU’s • First: task-specific cameras • Then: programmable cameras – FPGA programmable camera arrays Thanks to Jiwon Kim, Michael Cohen
- Slides: 36