Programming Project #4 (proj4)15-463: Computational Photography
Natasha Kholgade
Programming Project #4 (proj4)
For the following image taken in Smith Hall, a rectified one has been constructed that sets the camera to be
parallel to one of the walls, by rectifying the door in the center. The width of the door is assumed to be
around half the height (which is a reasonable estimate).
For the image below, the whiteboard was used to rectify the image with the assumption that the whiteboard height
is two-thirds its width.
The following three are panoramas generated from stitching; for a set of images, the final image was set to the image in the center was
found, and iteratively images on the sides were warped and blended into this final image. Since the homography induces the points in
one image to overlap with those of the other image, the blending was done by first calculating the range of this overlap, and then
having a ramp blending function where the ramp extended over the range of overlap.
For the images shown below, the detected interest points (for the first image), MOPS matches and matches after RANSAC
iterations 2, 50 and 500 are shown.
When MOPS features were used to automatically find the correspondences across images, the following panoramas were obtained. The MOPS
features with RANSAC provide a pretty good method to find correspondences. RANSAC was run for 1000 iterations.
To perform touring into the picture, the corners of the back plane and the vanishing point were marked. Intersections of the line segments
between the vanishing point and the corners of the back plane with the image boundaries were used to set up the 3D box that represented the
scene. Once this was done, the homographies required to rectify the five planes were determined, the different planes were warped according
to their respective homographies, and the result was textured mapped to the surface of the 3D box and projected in MATLAB. Following are some images
and various views of the 3D geometry.
