Descriptors in feature detection/extraction

Search IconIcon to open search

Source: http://medium.com/data-breach/introduction-to-feature-detection-and-matching-65e27179885d
Backlinks:  Bag of words

Descriptors

  • A description of the local appearance around each feature point (keypoint)
  • The descriptor encodes ‘interesting’ information from the image into numbers and act as an identifier (‘fingerprint’) to differentiate between features
  • The description should ideally be invariant to changes (such as illumination, translation, scale, in-plane rotation) so that the feature can be found again, even if the image is transformed
  • Typically: for each feature point, there is a descriptor vector

Classes of descriptors:

  • Local descriptor
    • represents the point’s local neighbourhood
  • Global descriptor
    • describes the whole image
    • generally not very robust—changes in parts of the image may cause the descriptor to fail

Some algorithms for feature detection/descriptor generation

  • SIFT (scale-invariant feature transform)
  • SURF (speeded up robust feature)
  • BRISK (binary robust invariant scalable keypoints)
  • BRIEF (binary robust independent elementary features)
  • ORB  (oriented FAST  and rotated BRIEF)

Source: http://en.wikipedia.org/wiki/Bag-of-words_model_in_computer_vision

Feature representation using feature descriptors

  • Descriptors: vectors that represent local patches in the image
  • A good descriptor should be able to handle transformations such as rotation, intensity change, scale, etc.
  • e.g. SIFT: Scale-invariant feature transform
    • converts a patch to a vector in 128 dimensions
    • image as a bag of words – image as a collection of SIFT vectors – image as a collection of sparse numerical vector