hog

menpo.feature.hog(image, *args, **kwargs)[source]

Extracts Histograms of Oriented Gradients (HOG) features from the input image.

Parameters:
  • pixels (Image or subclass or (C, X, Y, ..., Z) ndarray) – Either the image object itself or an array with the pixels. The first dimension is interpreted as channels. This means an N-dimensional image is represented by an N+1 dimensional array.
  • mode ({dense, sparse}, optional) –

    The sparse case refers to the traditional usage of HOGs, so predefined parameters values are used.

    The sparse case of dalaltriggs algorithm sets window_height = window_width = block_size and window_step_horizontal = window_step_vertical = cell_size.

    The sparse case of zhuramanan algorithm sets window_height = window_width = 3 * cell_size and window_step_horizontal = window_step_vertical = cell_size.

    In the dense case, the user can choose values for window_height, window_width, window_unit, window_step_vertical, window_step_horizontal, window_step_unit and padding to customize the HOG calculation.

  • window_height (float, optional) – Defines the height of the window. The metric unit is defined by window_unit.
  • window_width (float, optional) – Defines the width of the window. The metric unit is defined by window_unit.
  • window_unit ({blocks, pixels}, optional) – Defines the metric unit of the window_height and window_width parameters.
  • window_step_vertical (float, optional) – Defines the vertical step by which the window is moved, thus it controls the features’ density. The metric unit is defined by window_step_unit.
  • window_step_horizontal (float, optional) – Defines the horizontal step by which the window is moved, thus it controls the features’ density. The metric unit is defined by window_step_unit.
  • window_step_unit ({pixels, cells}, optional) – Defines the metric unit of the window_step_vertical and window_step_horizontal parameters.
  • padding (bool, optional) – If True, the output image is padded with zeros to match the input image’s size.
  • algorithm ({dalaltriggs, zhuramanan}, optional) – Specifies the algorithm used to compute HOGs. dalaltriggs is the implementation of [1] and zhuramanan is the implementation of [2].
  • cell_size (float, optional) – Defines the cell size in pixels. This value is set to both the width and height of the cell. This option is valid for both algorithms.
  • block_size (float, optional) – Defines the block size in cells. This value is set to both the width and height of the block. This option is valid only for the dalaltriggs algorithm.
  • num_bins (float, optional) – Defines the number of orientation histogram bins. This option is valid only for the dalaltriggs algorithm.
  • signed_gradient (bool, optional) – Flag that defines whether we use signed or unsigned gradient angles. This option is valid only for the dalaltriggs algorithm.
  • l2_norm_clip (float, optional) – Defines the clipping value of the gradients’ L2-norm. This option is valid only for the dalaltriggs algorithm.
  • verbose (bool, optional) – Flag to print HOG related information.
Returns:

hog (Image or subclass or (X, Y, ..., Z, K) ndarray) – The HOG features image. It has the same type as the input pixels. The output number of channels in the case of dalaltriggs is K = num_bins * block_size *block_size and K = 31 in the case of zhuramanan.
Raises:
  • ValueError – HOG features mode must be either dense or sparse
  • ValueError – Algorithm must be either dalaltriggs or zhuramanan
  • ValueError – Number of orientation bins must be > 0
  • ValueError – Cell size (in pixels) must be > 0
  • ValueError – Block size (in cells) must be > 0
  • ValueError – Value for L2-norm clipping must be > 0.0
  • ValueError – Window height must be >= block size and <= image height
  • ValueError – Window width must be >= block size and <= image width
  • ValueError – Window unit must be either pixels or blocks
  • ValueError – Horizontal window step must be > 0
  • ValueError – Vertical window step must be > 0
  • ValueError – Window step unit must be either pixels or cells

References

[1]N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection”, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005.
[2]X. Zhu, D. Ramanan. “Face detection, pose estimation and landmark localization in the wild”, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012.