Integral Channel Feature Descriptor

Figure 7.3: Images of the channels used by ICF. From left to right, the original image followed by the images of the different channels: 8 channels for the quantized gradient phase, 1 channel for the gradient magnitude, and 3 channels for the LUV components, respectively.

HOG variants use cells of varying shapes, and it was found that one way to speed up computation of the magnitude histograms was to use the integral image again. Thus, at the intersection of HOG and Haar features, Integral Channel Features have recently demonstrated interesting performance, effectively constituting a generalization of HOG that exploits the integral image.

The feature values that can be extracted are obtained by summing areas computed not directly from the original image but from different secondary images, obtained through nonlinear processing of the region being characterized. The most common possible transformations are the gradient-phase channels already seen in HOG, the gradient-magnitude image, the grayscale image itself, and, when available, two additional channels representing chrominance. The gradient is often computed using Sobel, but several experiments show that a simple derivative filter produces satisfactory results as well. Here too, the Sobel phase can be used with or without its sign, depending on the specific application.

The scalar representing the feature to be extracted is simply the sum over a rectangular area within one of the computed channels.