SURF: Speeded-Up Robust Features CPSC 643, Presentation 2 SURF: Speeded-Up Robust Features Herbert Baya, Andreas Essa, Tinne Tuytellaresb, Luc Van Goola,b aETH Zurich bK. U. Leuven, ESAT-PSI Sternwartstrasse 7 Kasteel Arenberg 10 CH-8092 Zurich B-3001 Leuven Switzerland Belgium Computer Vision and Image Understanding (CVIU) Vol. 110, No. 3, pp. 346-359, 2008.

Mostly Related Works Harris Corner Detector - Harris 1988 Laplacian of Gaussian - Lindeberg 1998 Harris - Laplace Detector - Mikolajczyk 2001 Difference of Gaussian - Lowe 2004

Mostly Related Works Harris Corner Detector - Harris 1988 Laplacian of Gaussian - Lindeberg 1998 Harris - Laplace Detector - Mikolajczyk 2001 Difference of Gaussian - Lowe 2004

Mostly Related Works Harris Corner Detector - Harris 1988 Laplacian of Gaussian - Lindeberg 1998 Harris - Laplace Detector - Mikolajczyk 2001 Difference of Gaussian - Lowe 2004

Mostly Related Works Harris Corner Detector - Harris 1988 Laplacian of Gaussian - Lindeberg 1998 Harris - Laplace Detector - Mikolajczyk 2001 Difference of Gaussian - Lowe 2004

Mostly Related Works Harris Corner Detector - Harris 1988 Laplacian of Gaussian - Lindeberg 1998 Harris - Laplace Detector - Mikolajczyk 2001 Difference of Gaussian - Lowe 2004

Other Related Works Salient Region Detector - Kadir 2001 Edge-based Region Detector - Jurie 2004

Motivation Using Laplacian of Gaussian, one could obtain scale invariant features. Lowe uses difference of Gaussian to approximate Laplacian of Gaussian. This paper uses Hessian - Laplacian to approximate Laplacian of Gaussian, to improve calculation speed.

Methodology Using integral images for major speed up Integral Image (summed area tables) is an intermediate representation for the image and contains the sum of gray scale pixel values of image.

Detection Hessian-based interest point localization Lxx(x,y,σ) is the Laplacian of Gaussian of the image. It is the convolution of the Gaussian second order derivative with the image. Lindeberg showed Gaussian function is optimal for scale-space analysis. This paper use Dxx to approximate Lxx.

Detection Approximated second order derivatives with box filters. Scale analysis with constant image size

Description Orientation Assignment x response y response Side length = 4s Cost 6 operation to compute the response Circular neighborhood of radius 6s around the interest point (s = the scale at which the point was detected)

Description Dominant orientation The Haar wavelet responses are represented as vectors Sum all responses within a sliding orientation window covering an angle of 60 degree The two summed response yield a new vector The longest vector is the dominant orientation

Description Split the interest region (20s x 20s) up into 4 x 4 square sub-regions. Calculate Haar wavelet response dx and dy and weight the response with a Gaussian kernel. Sum the response over each sub-region for dx and dy, then sum the absolute value of resp- onse. Normalize the vector into unit length

Matching Fast indexing through the sign of the Laplacian for the underlying interest point The sign of trace of the Hessian matrix Trace = Lxx + Lyy

Experiments

Experiments

Experiments

Analysis and Conclusion SURF is faster than SIFT by 3 times, and has recall precision not worse than SIFT. SURF is good at handling image with blurring or rotation. SURF is poor at handling image with viewpoint or illumination change.