Zhengyou Zhang: SFM Example

SFM Example

Image pair with points of interest overlayed

This is a difficult set of data because the scene is quite flat except at the upper right corner where there is a big rock. You can get this pair (first (222k) and second (219k) images) in PGM. The two images are calibrated. Their camera projection matrices are given here and there. The extrinsic parameters are of course irrelevant. Therefore, you can equivalently use the intrinsic matrices for the first and second images.

We first use Plessey-Harris detector to extract two sets of points (first (7k) and second (7k) point sets). There are 1131 and 1137 points, respectively. On the left, we show the image pair with points overlayed. You can click here (320k) to get a big picture.


Point matches found by a correlation technique

The correlation technique, as implemented in the software IMAGE-MATCHING, has established 624 point matches. You can get the coordinates of these point matches by clicking here (4k) and there (4k).

On the left, we show the image pair with point matches overlayed. You can click here (335k) to get a big picture.


Point matches found by a relaxation technique

The relaxation technique, as implemented in the software IMAGE-MATCHING, has established 697 point matches. You can get the coordinates of these point matches by clicking here (4k) and there (4k).

On the left, we show the image pair with point matches overlayed. You can click here (332k) to get a big picture.


Point matches after outlier detection based on Least-Median-Squares

The robust technique based on the least-median-square (LMedS) method, as implemented in the software IMAGE-MATCHING, has detected 57 outliers. That is, 640 point matches are considered as good matches. You can get the coordinates of these point matches by clicking here (4k) and there (4k).

On the left, we show the image pair with point matches overlayed. You can click here (335k) to get a big picture.


Motion and structure from motion

Using our software SFM, we compute the Motion.txt between the two images and the 3D reconstruction of the scene.

Here's a VRML based scene model (26K) of the reconstructed scene after Delaunay triangulation. When vrweb is used, set Display option to Texturing and turn off Lighting Calculations. You can download the texture image for this model.

What follows are two views of this reconstruction. If you are able to do cross-eye fusion, you should perceive the 3D structure.

3D reconstruction represented by a pseudo stereogram

The above two pictures are also available in big format. Click here (19k) for the left one and there (19k) for the right one.


<=Go to Zhengyou Zhang home page