Auto-Stitching Photo Mosaics

Sean Chen

Part 1: IMAGE WARPING and MOSAICING

Image rectification

.
Input
.
with hand labled correspondences
.
Rectified
.
Input
.
with hand labled correspondences
.
Rectified

Image mosaics

.
with hand labled correspondences
.
with hand labled correspondences
.
Second image warped to the perspective of the first image, and added together with alpha=0.5 where the images overlap. Some edge artifacts.
.
Now we blend with alpha feathering. for each image, each pixel's alpha is porpotional to how far away it is from the border of the image.

.
.
.

.
.
.

Part 2: FEATURE MATCHING for AUTOSTITCHING

. .
Used Harris Interest Point Detector to detect corners in the image.

. .
Use Adaptive Non-maximum Suppression (ANMS) to get a nice even spread of corner points throughout the image, so it's nice for the homography computation.

. .
Extracted Feature Descriptors, which are 8x8 blurred, black and white patches around each corner point. Then, use pairwise Euclidian distance to figure out which features match between the left and right images.

. .
Since least squares is very sensitive to outliers (the points in red), we remove them with RANSAC.

.
Compute homography, use inverse warping to warp the right image to the left, and then use alpha feathering (closer to center = higher alpha) to blend them together.

Side by side results: hand labeled vs. auto stitched

. .
. .
. .