Project2: Features

Project 2: Features [By Pakawat Phalitnonkiat (pp287) and Rina Tse (rt297)]

Design and Performance

The performance is quite good but not good as SIFT. We have implemented the function to compute Harris score by filtering the image to find gradients in x- and y-axes. Then we put them in the Harris matrix H and compute the maximum eigenvalue and corresponding eigenvector. However, the eigenvector direction could be defined in 2 ways. We use arctan to find the angle of the eigenvector but the arctan range is [-pi/2, pi/2], so we have an option to add it with pi if the gradient in x direction is negative. So, we ended up with the eigenvector with direction pointing to the brighter area.

We only collect about top 2% of the Harris scores to be features. That is, we set the threshold to be at z-score of 2.05 and normalized all the features. This turns out to be a good idea since we don't have to worry about the size of the images which could produce too many features and take more time to compute.

1. ROC

1.1 Graf

· ROC of graf1 vs graf2

$Description: Description: I:\final edit\Figs\graf12.png$

· ROC of graf1 vs graf4 (NOTE: no sift results provided)

$Description: Description: I:\final edit\Figs\graf14.png$

1.2. Yosemite

· ROC of Yosemite 1 and 2.

$Description: Description: I:\final edit\Figs\y12.png$

2. Images of Harris operator on Graf and Yosemite

Graf:

· Graf1

$Description: Description: Description: Description: C:\Users\CAM\Downloads\submit\webpage\GrafHarris.jpg$

· Graf2

$Description: Description: I:\final edit\raw_Kun\graf2Harris.jpg$

· Graf4

$Description: Description: I:\final edit\raw_Kun\harrisgraf4.jpg$

Yosemite:

· Yosemite 1

$Description: Description: Description: Description: C:\Users\CAM\Downloads\submit\webpage\YoHarris.jpg$

· Yosemite 2

$Description: Description: I:\final edit\raw_Kun\yo2harris.jpg$

3. AUC

Descriptor	Matcher	Bikes	Wall	Leuven	Graf	Yosemite
Simple	SSD	0.691	0.557	0.594	0.289901	0.179564
Simple	Ratio	0.529	0.607	0.581	0.615195	0.875559
MOPS	SSD	0.233	0.250	0.265	0.301343	0.152834
MOPS	Ratio	0.890	0.750	0.796	0.775898	0.937792
	Average	0.566	0.541	0.559	0.495584	0.5364373

Strengths & Weaknesses

Strength:

· Harris score is stable in most situations.

· We also implemented the Harris detector so that the threshold varies with the scale in intensity of the image.

· MOPS also provides us the invariance in rotation and illumination.

Weakness:

· The simple descriptor is not invariant.

· MOPS is still not good enough at rotation.

· The descriptors we implemented are not good as SIFT.

Other Images

We have used our own images to test our program. Here are the results:

The results came out to be desirable. It matches correctly about 75-90%.

· Testing the scaling images. Even though our descriptor is not scale invariant but it turned out to good in this example.

$Description: Description: Description: Description: C:\Users\CAM\Downloads\submit\Custom Img\match custom imgs\result 1 MOPS.png$

· Testing images in different illumination.

$Description: Description: Description: Description: C:\Users\CAM\Downloads\submit\Custom Img\match custom imgs\Result 2 MOPS.png$

· Testing scaling and rotating images.

$Description: Description: Description: Description: C:\Users\CAM\Downloads\submit\Custom Img\match custom imgs\result 4 MOPS.png$

Extra Credit Items

· Adaptive Non-Maximal Suppression

This is implemented in adaptiveSuppress() and can be enabled by calling it (uncomment the line) in ComputeHarrisFeatures()

Parameters: cRobust = 0.9 as suggested in the paper. Number of features nip as shown below:

~900 features

2236 features

Without Adaptive Non-Maximal Suppression With Adaptive Non-Maximal Suppression

· Alternative Matching score

We tried using (1-cosine similarity) to measure the distance between two features instead of using the Euclidean distance. It was found that the AUC results using (1-cosine similarity) improve over the simple Euclidean distance results but do not improve over the ratio test results.