Daniel Puckowski
In this paper, I present a novel quasi-rotation invariant interest point descriptor, coined COIF (Concentric Oval Intensity Features). The descriptor is straightforward to implement and feature matching is time efficient. COIF may be used to detect rotated images and may be used for image stitching in panorama applications. COIF demonstrates the feasibility of using luminance histograms for feature matching.
| Description | SIFT | COIF |
|---|---|---|
| Instances Equal | 55 | 55 |
| SIFT Better | 11 | - |
| COIF Better | - | 8 |
| Accuracy (%) | 98.9589 | 98.5205 |
| More Accurate (%) | +0.4384 | - |
| Accuracy Range | Count |
|---|---|
| 100% | 60 |
| 99-95% | 6 |
| 94-90% | 4 |
| 89-85% | 0 |
| 84-80% | 3 |
| Accuracy Range | Count |
|---|---|
| 100% | 65 |
| 99-95% | 4 |
| 94-90% | 1 |
| 89-85% | 2 |
| 84-80% | 0 |
| 79-75% | 1 |
| Dataset | COIFv6 Success | COIFv6 Failure | SIFT Success | SIFT Failure | COIFv6 vs. SIFT |
|---|---|---|---|---|---|
| SPW Dataset (2020) | 88.373% | 11.627% | 95.455% | 4.545% | -7.082% |
| Dataset for Stitching with Multiple Registrations (2018) | 65.286% | 35.714% | 50.000% | 50.000% | +15.286% |
| VPG Dataset (2020) | 90.910% | 9.090% | 44.455% | 54.545% | +46.455% |
| Effect | Accuracy Range |
|---|---|
| Light Variation | +/- 10% |
| Perspective Transformation | 25% |
| Scale Change | +/- 20% |
| Guassian Blur | +3 pixel radius |
| Average Matching Time | Median Matching Time | Image Pair Count | Pixels Processed Count |
|---|---|---|---|
| 4,283 milliseconds | 1,985 milliseconds | 45 | 25,543,680 |
| Average Matching Time | Median Matching Time | Image Pair Count | Pixels Processed Count |
|---|---|---|---|
| 3,778 milliseconds | 1,577 milliseconds | 45 | 25,543,680 |
Matching times include time to identify corners, time to generate descriptors, and time for feature matching.
| Bin Merge Count | Number of Times Used | Percent Occurrence |
|---|---|---|
| 1 | 38 | 69.09% |
| 2 | 3 | 5.45% |
| 3 | 4 | 7.27% |
| 4 | 5 | 9.09% |
| 5 | 5 | 9.09% |
| Bin Merge Count | Iteration | Count | Percent Occurrence |
|---|---|---|---|
| 1 | 1 | 66 | 51.96% |
| 2 | 1 | 16 | 12.59% |
| 3 | 1 | 11 | 8.66% |
| 4 | 1 | 11 | 8.66% |
| 4 | 2 | 1 | 0.78% |
| 4 | 6 | 1 | 0.78% |
| 5 | 1 | 7 | 5.51% |
| 5 | 2 | 2 | 1.57% |
| 5 | 4 | 1 | 0.78% |
| 5 | 5 | 1 | 0.78% |
| 5 | 7 | 2 | 1.57% |
| 5 | 8 | 1 | 0.78% |
| 5 | 9 | 7 | 5.51% |
Given the test image pair set, 51.96% of all image pairs yielded passing feature matches with the default COIFv6 parameters. Given the test image pair set, 87.38% of all image pairs yielded passing feature matches within the first 5 iterations.