Panoramic test data for ISPRS WG V-3

Background

The test data comprises sets of panoramic images (i.e., each having 360 degs. horizontal field of view), and the resulting 3D point distributions and models. Each 3D point distribution is recovered by applying omnidirectional multibaseline stereo [1] on the multiple panoramic images. This has the advantage of being able to extract 3D information of a wide scene directly. In addition, by virtue of the wide view, the intermediate process of recovering camera motion (using the 8-point algorithm in our case) is stable.

Each panoramic image is created by following these steps:

1. Recover intrinsic camera parameters (focal length, radial distortion factor, aspect ratio) through a separate calibration step;
2. Take sequence of images while rotating camera about a vertical axis (the axis has to pass through the camera optic center to avoid depth parallax);
3. Undistort images and project rectilinear images to surface of cylinder whose cross-sectional radius is the focal length;
4. Perform global (phase correlation) and local (iterative gradient descent) image registration to determine relative translation between successive images;
5. Blend all images using weighted bilinear interpolation [6,7].

Steps 1-3 can be replaced by iteratively compositing and recomputing the focal length [3]. More details of the omnidirectional multibaseline stereo technique can be found here.

Test sets

• Test set #1 (synthetic room):
  • 4 panoramic images of a synthetic room (PPM files, 1.1M each):
    • Pan0 (reference),
    • Pan1,
    • Pan2,
    • Pan3.
  • Correct 3D point distribution:
    • VRML format (635K),
  • 3D distribution from multibaseline stereo:
    • VRML format (53K),
  • Extracted (spatially filtered and resampled, NO texture-map) model from stereo data:
    • Unfiltered mesh (VRML, 210K),
    • Filtered mesh (VRML, 210K).
• Test set #2 (CRL Vision-Based Interaction Lab):
  • 6 panoramic images of Cambridge Research Lab's Vision-based Interaction Lab. (PGM files, 572K each):
    • Pan0,
    • Pan1,
    • Pan2,
    • Pan3,
    • Pan4,
    • Pan5.
  • Two sets of 3D point distribution from multibaseline stereo:
    • Pan2 as reference (VRML, 43K),
    • Pan3 as reference (VRML, 42K).
  • Extracted (spatially filtered) texture-mapped model:
    • VRML format (1.22M)

Relevant technical reports/papers

1. S.B. Kang and R. Szeliski, "3-D scene data recovery using omnidirectional multibaseline stereo," Conf. on Computer Vision and Pattern Recognition, June 1996, San Francisco, CA, pp. 364-370 (To appear in International Journal of Computer Vision, 1997. Also as Tech. Rep. CRL 95/6, Digital Equipment Corporation, Cambridge Research Lab, Oct. 1995).
2. A. Johnson and S.B. Kang, Registration and Integration of Textured 3-D Data, Tech. Rep. CRL 96/4, Digital Equipment Corporation, Cambridge Research Lab, Oct. 1996.
3. S.B. Kang and R. Weiss, Characterization of Errors in Compositing Panoramic Images, Tech. Rep. CRL 96/2, Digital Equipment Corporation, Cambridge Research Lab, June 1996.
4. R. Szeliski, S.B. Kang, and H.-Y. Shum, "A parallel feature tracker for extended image sequences," IEEE Int'l Symposium on Computer Vision, Coral Gables, FL, Nov. 1995, pp. 241-246 (To appear in Computer Vision and Image Understanding, 1997).
5. S.B. Kang, A. Johnson, and R. Szeliski, Extracting Concise and Realistic 3-D Models from Real Data, Tech. Rep. CRL 95/7, Digital Equipment Corporation, Cambridge Research Lab, Oct. 1995.
6. R. Szeliski and S.B. Kang, "Direct methods for visual scene reconstruction," IEEE Workshop on Representations of Visual Scenes, Cambridge, MA, June 1995, pp. 26-33.
7. R. Szeliski, "Image Mosaicing for Tele-Reality Applications," IEEE Workshop on Applications of Computer Vision, Sarasota, FL, December 1994, pp. 44-53.
8. R. Szeliski and J. Coughlan, "Hierarchical Spline-Based Image Registration," IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Seattle, WA, June 1994, pp. 194-201.

Acknowledgement

Andrew Johnson wrote the program to create the simplified texture-mapped 3D mesh.

Last modified: Tue Feb 4 14:11:12 EST 1997 by Sing Bing Kang.
Feedback is welcome.