Brian Zhu (brian_zhu@berkeley.edu)

Feature Finding

Original Image:

Harris Corners:

ANMS Corners (choosing top 50):

Feature Descriptors:

Feature Matching

Original Images:

ANMS Points (top 500):

Matching Points:

Mosaics

I decided to give a shot at stitching the same pictures used in project 4A

Balcony

Original Images:

Old Mosaic:

New Mosaic:

Target

Original Images:

Old Mosaic:

New Mosaic:

Hollywood

Original Images:

Old Mosaic:

New Mosaic:

Conclusions

Truthfully speaking, the new auto-stitched panoramas don't look significantly better than the old ones from the previous project. However, in order to achieve a decent mosaic in project 4a, it took a lot of painstaking tries to meticulously plot correspondences. Here, the correspondences are automatically found, and the algorithm can easily find many more as well (e.g. 50 pairs of points instead of 8). In addition, it also is possible to use different sets of correspondences between each pair of images, which previously led to some nasty inaccuracies when manually stitching. So while it may not be very apparent in the final result, the speed and accuracy of automatically finding correspondences makes this algorithm extremely powerful when generating mosaics.