Class DescribePointSurf<II extends ImageGray<II>>

Direct Known Subclasses:

public class DescribePointSurf<II extends ImageGray<II>> extends Object

Implementation of the SURF feature descriptor, see [1]. SURF features are invariant to illumination, scale, and orientation. Both the orientated and unoriented varieties can be computed. SURF-64 describes an interest point using a 64 values that are computed from 16 sub regions. Each sub-region contributes 4 features, the sum of dx,|dx|,dy,|dy|, where dx and dy are the local derivatives.

To improve performance (stability and/or computational) there are a few (intentional) deviations from the original paper.

  • Haar wavelet or image derivative can be used.
  • Derivative sample coordinates are interpolated by rounding to the nearest integer.
  • Weighting function is applied to each sub region as a whole and not to each wavelet inside the sub region. This allows the weight to be precomputed once. Unlikely to degrade quality significantly.

Usage Notes:
If the input image is floating point then normalizing it will very slightly improves stability. Normalization in this situation means dividing the input image by the maximum pixel intensity value, typically 255. In stability benchmarks it slightly change the results, but not enough to justify the runtime performance hit.

[1] Bay, Herbert and Ess, Andreas and Tuytelaars, Tinne and Van Gool, Luc, "Speeded-Up Robust Features (SURF)" Comput. Vis. Image Underst., vol 110, issue 3, 2008

  • Field Details

    • widthLargeGrid

      protected int widthLargeGrid
    • widthSubRegion

      protected int widthSubRegion
    • widthSample

      protected double widthSample
    • weightSigma

      protected double weightSigma
    • useHaar

      protected boolean useHaar
    • inputType

      protected Class<II extends ImageGray<II>> inputType
    • featureDOF

      protected int featureDOF
    • ii

      protected II extends ImageGray<II> ii
    • weight

      protected Kernel2D_F64 weight
    • gradient

      protected SparseScaleGradient<II extends ImageGray<II>,?> gradient
    • gradientSafe

      protected SparseImageGradient<II extends ImageGray<II>,?> gradientSafe
    • radiusDescriptor

      protected int radiusDescriptor
  • Constructor Details

    • DescribePointSurf

      public DescribePointSurf(int widthLargeGrid, int widthSubRegion, double widthSample, double weightSigma, boolean useHaar, Class<II> inputType)
      Creates a SURF descriptor of arbitrary dimension by changing how the local region is sampled.
      widthLargeGrid - Number of sub-regions wide the large grid is. Typically 4
      widthSubRegion - Number of sample points wide a sub-region is. Typically 5
      widthSample - The width of a sample point. Typically 4
      weightSigma - Weighting factor's sigma. Try 3.8
      useHaar - If true the Haar wavelet will be used (what was used in [1]), false means an image gradient approximation will be used. False is recommended.
    • DescribePointSurf

      public DescribePointSurf(Class<II> inputType)
      Create a SURF-64 descriptor. See [1] for details.
  • Method Details

    • createDescription

      public TupleDesc_F64 createDescription()
    • setImage

      public void setImage(II integralImage)
    • describe

      public void describe(double x, double y, double angle, double scale, boolean normalize, TupleDesc_F64 ret)

      Computes the SURF descriptor for the specified interest point. If the feature goes outside of the image border (including convolution kernels) then null is returned.

      x - Location of interest point.
      y - Location of interest point.
      angle - The angle the feature is pointing at in radians.
      scale - Scale of the interest point.
      normalize - Apply L2 normalization to the descriptor
      ret - storage for the feature. Must have 64 values.
    • features

      public void features(double c_x, double c_y, double c, double s, double scale, SparseImageGradient gradient, double[] features)

      Computes features in the SURF descriptor.

      Deviation from paper:

      • Weighting function is applied to each sub region as a whole and not to each wavelet inside the sub region. This allows the weight to be precomputed once. Unlikely to degrade quality significantly.

      c_x - Center of the feature x-coordinate.
      c_y - Center of the feature y-coordinate.
      c - cosine of the orientation
      s - sine of the orientation
      scale - The scale of the wavelets.
      features - Where the features are written to. Must be 4*(widthLargeGrid*widthSubRegion)^2 large.
    • getDescriptionLength

      public int getDescriptionLength()
    • getInputType

      public Class<II> getInputType()
    • getCanonicalWidth

      public int getCanonicalWidth()
      Width of sampled region when sampling is aligned with image pixels
      width of descriptor sample
    • copy

      public DescribePointSurf<II> copy()
      Creates a new instance with the same configuration