Features

We support different types of visual features, which can be used to create the Bag of Features databases.

SIFT

vsearch.sift.calculate_sift(image, roi=None, only_keypoints=False)

Calculate SIFT descriptors and/or keypoints

Parameters:
  • image (array_like) – The input image array
  • roi (array_like) – An optional region of interest encoded as (x, y, width, height) If None, then the whole image is used
  • only_keypoints (bool) – If True then only keypoints will be computed
Returns:

  • descriptors (array_like) – NxD array containing the N D-dimensional descriptor vectors
  • keypoints (list) – List of N cv2.Keypoint objects
vsearch.sift.sift_file_for_image(path)

Return SIFT descriptor filename corresponding to an image path

Color names

vsearch.colornames.calculate_colornames(image, roi=None, keypoints=None)

Calculate colornames descriptors and/or keypoints

For color names we still use SIFT to detect keypoints. For each keypoint we use a radius that is twice that of the corresponding SIFT feature, and calculate a histogram of color name probabilities within this region. The histogram is then normalized to not depend on the chosen radius.

Parameters:
  • image (array_like) – The input image array
  • roi (array_like) – An optional region of interest encoded as (x, y, width, height) If None, then the whole image is used
  • keypoints (list) – List of keypoints for which to compute the color names descriptors. If None, then SIFT keypoints for the image will be computed.
Returns:

  • descriptors (array_like) – NxD array containing the N D-dimensional descriptor vectors
  • keypoints (list) – List of N cv2.Keypoint objects
vsearch.colornames.cname_file_for_image(path)

Return color names descriptor filename corresponding to an image path

vsearch.colornames.colornames_image(image, mode='index')

Apply color names to an image

Parameters:
  • image (array_like) – The input image array (RxC)
  • mode (str) – If ‘index’ then it returns an image where each element is the corresponding color name label. If ‘probability’, then the returned image has size RxCx11 where the last dimension are the probabilities for each color label. The corresponding human readable name of each label is found in the COLOR_NAMES list.
Returns:
Return type:

Color names encoded image, as explained by the mode parameter.