40 corresponsences are defined mainly on the zebra lines and windows:
pic 1 after warping
example 1: Bicycle
example 2: Billboard
By simply taking average at overlapping pixels, I get the following result.
I set the threshold on the ratio between the first and the second nearest neighbors to 0.5 and applied reciprocal matching. The results are as follow.
There are a few outliers.
To remove outliers, I applied RANSAC and set epsilon to 6, number of loops to 600. The results are as follow.
Using the homogrphy determined by the inlier group, I get the following warping result:
Here is the comparison of automatically stitched mosaic using the above warping result and manually stitched mosaic from part 1:
The result of manually stitched mosaic is slightly better than the automatically stiched one. This is probably because the corresponding keypoints manaully selected are more evenly distributed than the ones filterd by RANSAC. As we can see in previous part, most points on the zebra stripes are filtered out by RANSAC.