Proj5 Part1 IMAGE WARPING and MOSAICING

Recover Homographies

The least squre problem can be viewed in this form. But sovling this we obtain the transform matrix H.

Warp the Images

After obtaining the transform matrix, we can use the similar method as in project3 to warp the pictures. By the mean time, I record the offsets from the unwarped picture for later procedures. Here are the example of rectified pictures

Blend the images into a mosaic

I blend the pictures one after another. And use the offset to track the position. At first I simply set alpha=0.5 But then I realize as the number of pictures increased, the first picture will be dimmer and dimmer.

Instead of use a mask, I want to try find the intersection and adjust it. So I find the intersection between the polygon of the two pictures to be blended. And divide the value of this area by 2. Other area just keep the original value. Since I blend the pictures by simply copy the value of the whole picture, the calculated intersection does not 100% represent the real intersetction. So we can see some darker area with straight edges. So use a mask to adjust the alpha channel will be a better and faster way. But I still choose to show this result. Original pictures:

Coarsed result:

After adjusting:

What I learn

The point correspondance matters a lot. If I only mark point in just a certain area, the result is not good. So I need to spread the point evenly and include all the importan features. For better result, each picture need more than 15 points.