Real-Time Image Mosaicing

1
Real-Time Image Mosaicing

• --Research Project for 15-869, Nov 29, 1999
• Li Zhang

2
Image Mosaicing

• Stitching together images taken from different
  viewpoints seamlessly
• Representation cylindrical, spherical, and
  perspective panoramas

3
Problem Statement

• Given two images I0 and I1 of a scene, find a
  transform T that minimizes

4
Special cases

• If the scene is planar, and I0 and I1 are taken
  from different viewpoints with different
  orientation,
• If the scene is arbitrary, and I0 and I1 are
  taken from same viewpoint with different
  orientation,
• T is a planar perspective transform M

5
My work

• Given two images I0 and I1 of a scene, find a
  transform M that minimizes
• Search M interactively in real-time

6
Illustration
7
Algorithm Details(1)

• Given I0, I1 and current M, we iteratively update
  M using
• We wish D to minimize

8
Algorithm Details(2)

• We warp I1 into I1 such that
• We instead wish D to minimize

9
Acceleration Details(1)

• We find d by solving a least-square minimization
  problem with following acceleration techniques
• Approximate the Jacobian as

10
Acceleration Details(2)

• Warp I1 from I1 using hardware-accelerated
  texture mapping
• Perform minimization hierarchically using OpenGL
  Mipmaps.

11
Results(1)

• Real images of University Center Mural

12
Results(2)

• Full view panorama of UC Mural

13
Summary(1)

• Good features
• Users can interactively create image mosaics in
  real-time
• The transform is locally optimized

14
Summary(2)

1. 8-parameter transform sometimes get stuck in
   local minimum
2. The size of image is limited by frame buffer

15
Reference

1. R. Szeliski and H.-Y. Shum. Creating full view
   panoramic image mosaics and texture-mapped
   models. Computer Graphics (SIGGRAPH'97), pages
   251-258, August 1997.