Speeded up robust feature(SURF) ?

Hi,

SURF uses scale-space created by applying kernels of increasing size to the original image, to make it scale invariant.

Now since each level of the space has, image with their own feature point. I'm wandering how SURF determine which level of scale it has to take to determine the dominant orientation of the feature point to be rotation invariant?

Answer 1

[David Young]

Not really a MATLAB question - but anyway, the scale at which the feature was detected is used to determine its dominant orientation. See the second sentence of section 4.1 of the paper by Bay et al.

[Kyle]

>.< sorry. yeah this is not a matlab question but decided to post the question here anyway cause many professor here, and i dont know any forum suitable for me to ask this question.

From that journal, the way i interpret is that at each level of scale space, the dominant orientation of of each interest point is determine with the sum of response of Haar wavelet transform.

But then if this is the case, then i cant image how it is going to match those point across so many scale levels

[David Young]

I don't fully understand your last point. Each SURF feature is detected at a specific scale (by non-maximum suppression in scale-space). To get a descriptor for that feature, the dominant orientation at the feature's scale is found, and a patch of the image is reoriented with respect to the that direction, before the responses are measured and put into the description vector. This means that two features at different scales and orientations can have the same descriptor, and so will match because there's no information about the original scale or orientation in the descriptor.

[Kyle]

Thanks, i can see the overall picture now. However there still are few missing link on my understanding.

1)A window(patch) is form around the interest point, then the descriptor vector is calculated according the the dominant orientation. So how come the correspondent matching point from other image with different scale can get same descriptor?

2)non-maximum suppression in scale-space,detects same feature across the scale space. lets say i have 2 image with different scale. with the non-maximum suppression i could detect a same feature across the scale space. How is this information used in matching?

Sorry for the trouble, i am quite new in this field.

[David Young]

1) Say image 1 is just like image 2 but expanded to twice the size. The scale of a feature in image 1 will be twice the scale of the corresponding feature in image 2. The patch extracted from image 1 will therefore be twice as big as the patch extracted from image 2, and all the filters used to get the description values will be twice as big too. That means that you get (in principle, anyway) the same description for the two patches.

2) Information about the scale is not used in matching. The descriptor is independent of the scale, due to the effect outlined above. So a feature at scale 1 in image 1 can match a feature at scale 2 in image 2 - there's no information about the original scale in the descriptor.

The descriptor has information about the pattern of light and dark in the pixels around the feature point - but this information is independent (to a good approximation) of the scale or orientation of the pattern.

[Kyle]

I think my understanding problem lies in the descriptor itself. I dont really understand how the information it stores can help in matching of feature point.

[David Young]

The descriptor has information about the pattern of light and dark in the pixels around the feature point - but this information is independent (to a good approximation) of the scale or orientation of the pattern.