Nyk Lotocky - njl36
Computer Vision
Project 1 - February 20, 2011

• My feature descriptor attempts to combine the simple descriptor and the SIFT descriptor. Rather than computing entire histograms of gradients, the descriptor is just the gradient at every point in a 9x9 window around the point. First, the feature is rotated to horizontal, similar to the MOPS descriptor, and the patch is convolved by the Sobel operator. The gradient orientation is then found at each point in the rotated patch.
• One of the design choices I had made was whether or not to include the gradient magnitudes as well. I ran some tests, and although in some cases having the magnitude increased the performance of the descriptor, in cases such as decreased intensity or blurryness, having simply the gradient did better. The reason the gradient is useful is because it captures the local shape of the feature in a concise manner, that is invariant to orientation.
• • ROC curves

    Figure: ROC curves for graf dataset

    Figure: ROC curves for yosemite dataset

    Figure: $\Vert\nabla\Vert$ included in custom descriptor. Performs marginally better on the graf dataset.

    

  • Harris images

    Figure: Harris image for graf dataset

    Figure: Harris image for yosemite dataset

  • Average AUC

    Descriptor	Data Set	SSD	Ratio
    Simple	graf	0.528553	0.519850
    	leuven	0.320724	0.531751
    	bikes	0.506024	0.506763
    	wall	0.445466	0.554764
    MOPS	graf	0.592051	0.583551
    	leuven	0.557452	0.601892
    	bikes	0.530373	0.555378
    	wall	0.646460	0.604222
    Custom	graf	0.577507	0.555842
    	leuven	0.648175	0.616482
    	bikes	0.583707	0.529553
    	wall	0.599429	0.566069
    Custom with magnitudes	graf	0.519003	0.535563
    	leuven	0.414312	0.569530
    	bikes	0.487185	0.534610
    	wall	0.517033	0.564836

  • Threshold Curves

    Figure: Simple, graf	Figure: MOPS, graf	Figure: Custom, graf	Figure: SIFT, graf
    Figure: Simple, yosemite	Figure: MOPS, yosemite	Figure: Custom, yosemite	Figure: SIFT, yosemite

• For some reason SSD does better as a distance measure than does the ratio test. In addition, on datasets primarily based on translation would be better served by the magnitude implementation of my descriptor, however, the orientation-only version performs better in the general sense.

• Real-life images. The images on the left of each set are identical, while the images on the right switch between rows. The columns are differentiated by MOPS or the custom descriptor.

  Figure: Using MOPS	Figure: Using custom descriptor
  Figure: Using MOPS	Figure: Using custom descriptor