Project 6

Summary

For the first part of this project, I warped together photos using a homography transformation with manually selected points and then blended them to create panoramas. Images were shot with my Nikon D80 and an inverse warp was used to create the warped images before blending. Blending for this part was simple linear blending, no pyramids used.

For part 2, I implemented the specified feature descriptor computation given points selected from a harris corner detector. Points detected with the corner detector were selected for feature computation with Adaptive Non-Maximal Suppression, which picks the most dominant points in a region. The descriptor computed was an 8x8 patch of the surrounding pixels selected from a 40 x 40 downsampled region. The features were matched using a simple distance metric (the provided dist2 function) and results were thresholded using the ratio between the first and second nearest neighbors. I chose a threshold value of 0.5 for my implementaiton.

The autostitched images were unable to use as many input images in my system, as I added them one at a time and the feature points computed for the already merged image didn't overlap as much with the panorama's massive amounts of features. For speed reasons the autostitched images were computed from half-size source images.

The full scale images are very large, and an archive of all of the images for this project can be found here.

Results

Scene 1 - Pausch Bridge

Pausch Bridge | Full Size

Pausch Bridge - Uncropped | Full Size

Pausch Bridge - Autostitch | Full Size

Sources

DSC_1451

DSC_1453

DSC_1455

DSC_1456

DSC_1458

DSC_1460

Scene 2 - Schenley Park

Schenley Park | Full Size

Schenley Park - Uncropped | Full Size

Schenley Park - Autostitch | Full Size

Sources

DSC_1462

DSC_1464

DSC_1466

DSC_1468

DSC_1469

Scene 3 - CMU

CMU | Full Size

CMU - Uncropped | Full Size

CMU - Autostitch | Full Size
Note: This scene had issues with feature detection with the descriptor. As such only two source images were able to be stitched together.

Sources

DSC_1470

DSC_1472

DSC_1474

DSC_1476

Rectified Images

Bridge rectified to be mostly facing forward

The UC rectified to a flat plane

Feature Points

Harris Points for DSC_1468

Feature Points after ANMS (1000 points)

Bells and Whistles

Day and Night | Full Size

Wall Dragon

Portal in the Park

Daylight
This didn't come out very well in my opinion. It was hard to keep features to detect when you just had the bridge color to work with, and the blacks being used as an alpha channel didn't work as well towards the edge.