15862 – Computational
Photography
Assignment 4: Feature Matching and Automatic Photostiching
- Nisarg Vyas (nisarg
AT cmu DOT edu)
Section
1: Automatic Photostiching
The goal of the project is
to find feature points automatically in a set of images, learn the homography and stich the set of
images into a mosaic. I show the results
on 2 different set of images.
Image Set 1:
0): Input Images
1):
Harris Corner Detection with Adaptive Non-Maxima Suppression
The corner detection with
the Adaptive Non-maxima suppression gives us a uniform distribution of interest
points selected in the image, which is a desired effect. Although here, the
“artifact” of using that is, I am getting feature points in the region of very
less ‘interest’ (e.g. sky).
2) Description Vector and Feature-space Outlier Rejection
This step finds out the similarity between each
point’s descriptor vector in one image with all the
other points’ descriptor vectors in the other image. The ratio between the best
match and second best match is also taken into account. For a pair of points to
be good feature point, the distance of the second best match should be atleast double than the distance of the best match.
3)
RANSAC (RAndom SAmple
Consensus) Estimate of homography
4)
Final Mosaic
Image
set 2:
0):
Input Images
1):
Harris Corner Detection with Adaptive Non-Maxima Suppression
2) Description Vector and Feature-space Outlier
Rejection
3)
RANSAC (RAndom SAmple
Consensus) Estimate of homography
4)
Final Mosaic
Section
2: Multi-scale processing
For this, I have followed
Lowe’s paper titled “Distinctive image features from scale-invariant keypoints”. The result that I show here is about the points
detected at different ‘charecterisitc’scales.
Section
3: Automatic Panorama Recognition
The
goal of this task was to identify different panoramas given a set of images
which might or might not contain mosaics. If panoramas exist,
stich them using autostiching.
Image
set 1
Input
Images
Output
Panorama
Image
set 2:
Output
Panoramas
(Note:
Image sizes are rescaled in arbitrary size to fit the images into the webpage.
These are not the original image sizes)