These images are from the corridor of National Lab of Pattern Recognition, Chinese Academy of Science(I got my master degree from there.).
I almost stood in one position to take these images, just changing view angles. So there are very apparent changes in depth and lighting among these images.
The bright red points in the first image represent the correspondence points I used in registering the first and second images. And The bright red points in the third image represent the correspondence points I used in registering the second and third images. I did this part of work manually.
Based on the correspondence points, I used Matlab to compute the transformation matrix, and then used a C procedure to get the final result(Since I am not familiar with this new software environment, I can only solve this assignment by the simplest way, although I know it is a little ugly.). It will totally take several seconds.
In the middle of this final image, the mosaicing result is poor. I think it is because of no correspondence points taken from the region. So the transformation results of the three original images in that region can be not accurate. I think it is a shortcoming of this method. Szeliski's method can conquer this shortcoming.
In the other parts of this final image, the mosaicing result is very satisfying. And the change in lighting in this image is much more natural than among the three original images.
According to the above result and analysis, I increased some correspondence points between the second and the third original images, shown as the above images(The bright white points are the new correspondence points. In order to see these correspondence points more clearly, here I increased the brightness of these two images by Paintshop.).
Apparently, this new result is much better than the above one. It shows that the analysis above is reasonable.