Professional Master Program - UW
Computer Vision - CSEP 576
Project 2:  Panoramic Mosaic Stitching
Student: Marcelo De Barros

Summary

The warping method was very straightforward - I pretty much followed the instructions from the slides. For the Lucas-Kanade algorithm, I noticed that when calculating the X and Y gradients for a certain pixel P(x,y), if you take into account the two horizontal neighbor pixels (0.5*(P(x+1,y) - P(x-1,y))) and the two vertical neighbor pixels (0.5*(P(x,y+1) - P(x,y-1))) the results are much more accurate. When calculating the size of the bounding box, I had some trouble to get the right values because I first thought that the displacements p[0] and p[1] were relative to the immediate previous image, which is incorrect: they are relative to the very first image. In other works, they are cumulative. The alpha blending wasn't that difficult. The transformation matrix required some thought, and although I decided to translate X by width/2, I don't think that this is really necessary. Overall, very good experience, nice project.

Images

For the full-resolution image, click here. For the panorama in the 360 degrees viewer, click here.

For the full-resolution image, click here. For the panorama in the 360 degrees viewer, click here.

For the full-resolution image, click here. For the panorama in the 360 degrees viewer, click here.