ng 
te 
es 
ne 
ng 
NS 
id 
ng 
  
147 
4.1 Applications 
As an application, we can take some pictures of a real scene by any means, and carry out a virtual visit of 
the scene, as if a 3D model of the entire scene had been computed. The user could move around the virtual 
camera, and each image would be synthesized on demand. The matching phase is still quite long and needs 
to be computed off-line, but the reprojection could be done at near real-time rates (in our (non-optimized) 
implementation, we currently transfer more than 20,000 points per second). The images obtained are more 
realistic than any 3D model we could have computed. 
Another application is very high-rate compression : given a video signal representing successive views of a 
scene, we could extract some carefully selected reference views, and then transmit only these views (themselves 
possibly compressed) along with the 17 parameters of the trilinear relation describing each successive view. The 
compression rate grows with the size of the images to transmit, because the number of parameters remains 
fixed. For non-static scenes, we need to compute 17 parameters by solid independently-moving object. 
4.2 Open problems 
e We need to take into account more than 2 reference views. The problem is : 
— What do we need to compute ? Do we need to match every point in every image ? 
— How do we transfer a point ? Do we only use its coordinates in 2 reference views ? In which views ? 
How do we combine more than 2 views ? 
— Which views do we need to transfer in the case of compression ? That is, which views do carry 
information which couldn’t be found in the other views ? In which way ? To what extent ? 
e Which form of the trilinear relations do we need ? Do we get more accurate results if we use simultaneously 
the various forms of trilinearity constraints ? 
References 
[Fau 92] O. Faugeras. What can be seen in three dimensions with an uncalibrated stereo rig? In G. Sandini, 
editor, Proceedings of the 2nd European Conference on Computer Vision, Santa Margherita Ligure, 
Italy, pages 563-578. Springer-Verlag, May 1992. 
[Gei 92] D. Geiger, B. Ladendorf, and A. Yuille. Occlusions and binocular stereo. In G. Sandini, editor, 
Proceedings of the 2nd European Conference on Computer Vision, Santa Margherita Ligure, Italy, 
pages 425-433. Springer Verlag, 1992. 
[Har 94] R. Hartley. Lines and points in three views - an integrated approach. 1994. 
[Lav 94] S. Laveau and O.D. Faugeras. 3-D scene representation as a collection of images. In Proceedings of the 
12th International Conference on Pattern Recognition, Jerusalem, Israel, volume 1, pages 689-691, 
1994. 
[Lon 95] Q. Long. Bilinearities and trilinearities. 1995. 
[Moh 92] R. Mohr, L. Quan, F. Veillon, and B. Boufama. Relative 3D reconstruction using multiples uncali- 
brated images. Technical Report RT 84-I-IMAG LIFIA 12, LIFIA—IRIMAG, 1992. 
[Oht 85] Y. Ohta and T. Kanade. Stero by intra and inter-scanline search using dynamic programming. IEEE 
Transactions on PAMI, 7(2): 139-154, 1985. 
[Sha 94] A. Shashua. Trilinearity in visual recognition by alignment. In Jan-Olof Eklundh, editor, Proceedings 
of the 3rd European Conference on Computer Vision, Stockholm, Sweden, pages 479-484. Springer 
Verlag, May 1994. 
[Wer 94] Th. Werner. Rendering Real-World Objects without 3-D Model. Technical report, Czech Technical 
University, Dept. of Control, Faculty of Electrical Engineering, Czech Technical University, Karlovo 
nam. 13, 12135 Praha, Czech Republic, 1994. . 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995