Fall 2018

CS194-26 - Project 6a

[Auto] Stitching Photo Mosaic

Regina Ongowarsito - cs194-26-aeq

10/10/2018

0: Summary

Using correspondence points, we warp two different images of the same scene via recovered homographies to create panoramas. In further developments, we introduce an algorithm to automatically identify effective correspondence points.

1: Image Rectifying

Lower Sproul Scenery

Before
After, rectifying on MLK's columns on the right of the image

Can't Read The Lecture Slide

Before
After, rectifying on the lecture slide

2: Mosaicing

.
Before, picture 1
Before, picture 2
Results, Alpha Blending
Results, Laplacian Blending
Before, picture 1
Before, picture 2
Results, Alpha Blending
Results, Laplacian Blending
Before, picture 1
Before, picture 2
Results, Alpha Blending
Results, Laplacian Blending

3: Automated Feature Matching for Autostitching

Finding Correspondence Points

Harris Corners
Lower Sproul, scaled 0.25
Applying ANMS post-Harris Corners
Lower Sproul, scaled 0.25

Results

Roof of Esh, Revisited
Before, picture 1
Before, picture 2
Manual
Results, Manual, Alpha Blending -- Warping Picture 1 to Picture 2
Auto
Results, Auto, Alpha Blending -- Warping Picture 1 to Picture 2
Results, Auto, Alpha Blending -- Warping Picture 2 to Picture 1
Lower Sproul, Revisited
Before, picture 1
Before, picture 2
Manual
Results, Manual, Alpha Blending -- Warping Picture 1 to Picture 2
Auto
Results, Auto, Alpha Blending -- Warping Picture 2 to Picture 1
Results, Auto, Alpha Blending -- Warping Picture 1 to Picture 2
Starbucks
Before, picture 1
Before, picture 2
Results, Auto, Alpha Blending
Results, Auto, Alpha Blending - Zoomed in to the right half

4: Summary

I've learned that homographies can warp images very well, but is very sensitive to the correspondence points used. Additionally, blending can be optimized by prioritizing edges -- if anything there's situations where your homography warp might be mostly incorrect, but still produces a believable warp if the edges of the warps blend together in a believable way. Additionally, I found it interesting the algorithms Harris Corners, ANMS, and Ransac works really well in a lot of cases! Certainly far easier than matching points manually, as we did in part 1.