Panorama Artifact
CSE 455: Computer Vision
Project 2 Artifact
Ekaterina Nepomnyashchaya and Chantal Murthy
Test sequence
"Building Test"(full size, 360 viewer)
Sequence with Kaiden panorama head
"Waterfront" by Katia and Chantal, CSE 455 Winter 2012 (full size, 360 viewer)
Sequence taken by hand
"Physics Building" by Katia and Chantal, CSE 455 Winter 2012 (full size, 360 viewer)
Harris Features
Harris features for yosemite
Harris features for graf
ROC comparison of our code vs. SIFT
Briefly introduce what you're showing in this section.
ROC comparison for Yosemite - MOPS + SSD (red), MOPS + ratio test (green), SIFT + SSD (blue), SIFT + ratio test (purple)
ROC comparison for graf - MOPS + SSD (red), MOPS + ratio test (green), SIFT + SSD (blue), SIFT + ratio test (purple)
Notes
- We had very good results for images with distinct features such as the test image. The water panarama also turned out surprisingly well.
- The area under the ROC curve for our feature dection ratio method was greater than that for MOPS + SSD. This means that ratio was the better feature matching method. Ofcourse, the SIFT technique was far superior.
- It was very difficult to rotate the camera exactly around the same axis for the handheld pictures. We ended up with a lot of vertical shift and tilting of the camera. This resulted in ghosting because angles had changed too much between adjacent pictures.
- Putting together these panoramas required quite a bit of messing with the threshold value. We would sometimes get supposedly no translation between adjacent images, and to fix it had to experiment with threshold values.
- Our implementation of finding the minimum suppression radius for each pixel in the image (the extra credit) took an unreasonable amount of time (about 15 minutes). This would not be compleatly unusable for any larger of a panarama.
- It took forever to copy/paste into the pairlist.txt. Writing a script would have been a good investment.
Extra Credit: Adaptive Non Maximum Suppression
- Implemented non-adaptive maximum suppression by finding the minimum distance from each pixel in the image until the value at the pixel fell below 0.9*value at the pixel some radius away.
- The top 500 pixels with maximum "minimum suppression radii" were chosen to be significant Harris features
- The hope was to yield a wider spatial distribution of significant Harris features in the image.