What worked well:
The test sequence worked nicely with our feature detection and alignment algorithms.

What didn't:
In general, we had to account for eliminating features along image edges that were detected as a result of warping the image.

What worked well:
It was nice to have a tripod for stability and being able to shift the viewing angle consistently between pictures. Corrections for vertical drift worked well for these shots.

What didn't:
For the indoor sequence, being too close to the whiteboard (and not having enough doodles!) seems to have resulted in too few distinct features in that region.
For the outdoor panorama, some features didn't align quite right as it was slightly windy, and there was a raindrop or two on the camera lens, resulting in an overall blurriness. We are also impressively bad photographers.

What worked well:
Stitching worked surprisingly well, since our solution had to handle shifts in the camera's viewing angle and position as well as a smaller number of image frames.

What didn't:
Several frames ended up being misaligned, overlapping one another.

---

Graphs of true positive rates on the y-axis against false positive rates on the x-axis.

Comparing our detected features against the known features in both image sets, the MOPS descriptors perform fairly well, but doesn't match the accuracy of SIFT in either case. In computing match scores, the ratio test is generally bests a simple SSD.

Image representations of applying the Harris operator on the sample images.

Our feature descriptor implements contrast invariance by normalizing the pixel values within the feature. For comparison, we include stitchings done on two of the test images, one of which has a decreased contrast. Without contrast invariance, features are matched incorrectly, which results in poor alignment.