I created several mosaics with the method in the handout. I did a simple alpha blend with biquadratic weights that are 1 at the center of the image and 0 around the whole edge.
As for the feature detection, consult the following images:
Original Features 1 | Suppressed/Matched/Agreed Features 1 | Suppressed/Matched/Agreed Features 2 | Original Features 2 |
In order for two samples to match, they must both be the best partner of the other with less than 80% of the sq-distance of the other's second best partner (i.e. a Russian-granny coefficient of 0.8). This is relatively permissive, but RANSAC easily picks up the slack.
I implemented the method from the paper. I was able to get novel views for the provided painting. With every other scene, MATLAB would just save straight white images, and I was too tired and frsutrated to work around MATLAB being bad. In place of novel views, I have provided images which can be printed, cut, and folded to create a model of the scene!
Original | Novel View 1 | Novel View 2 | Novel View 3 |
Original | Fold-Up |