Feature Descriptor

My feature descriptor is an 11x11 window of pixel values. I chose a larger descriptor in order to capture more information than the small 5x5 window. In order to make the descriptor rotation invariant, the dominant orientation (first eigenvector) at the detected feature location is computed, and the window is rotated about this angle. The values within the descriptor are normalized by subtracting by the mean and dividing by the standard deviation. Normalization helps make the descriptor invariant to changes in illumination.

ROC Curves

Graf Dataset (image pair 1 and 2)

AUC	Window	My Feature
SSD	0.630978	0.725945
Ratio Test	0.642158	0.842437

Yosemite Dataset (image pair 1 and 2)

AUC	Window	My Feature
SSD	0.934689	0.913754
Ratio Test	0.876972	0.889359

Harris Operator

Graf

Yosemite

Benchmark Results, Discussion

Bikes (average of AUC results for 6 image pairs)

The bikes dataset contains blurry and sharp images of a scene. The results were poor for the blurry images in this dataset for two reasons: first, very few features were detected in the blurry images. Second, none of my feature descriptors accounted for scale or frequency content.

*note: threshold was set lower in order to compute enough features for this dataset

AUC	Window	My Feature
SSD	0.363918	0.323983
Ratio Test	0.54590	0.579931

Graf (average of AUC results for 6 image pairs)

The graf dataset contains 3D perspective motion. My feature performed about as well as the simple window for the images with the more extreme perspective projection.

AUC	Window	My Feature
SSD	0.619872	0.570875
Ratio Test	0.669282	0.608005

Leuven (average of AUC results for 6 image pairs)

The leuven dataset contains brightness variation, but little other camera motion. My feature performed well on this dataset compared to the basic window due to the normalization.

AUC	Window	My Feature
SSD	0.47296	0.783543
Ratio Test	0.632996	0.664588

Wall (average of AUC results for 6 image pairs)

The wall dataset contains natural repetitive structure (a brick wall) and some camera motion. My feature performed better than the basic window, possibly because it was larger, and rotation invariant.

AUC	Window	My Feature
SSD	0.569964	0.687632
Ratio Test	0.658911	0.795984

Strengths and Weaknesses

The algorithms implemented for this project were very simple, but they were very sensitive to parameters and thresholds.

My Image – Example Matching Results