Integral Channel Descriptor

Figure 6.3: Image of the channels used by ICF. From left, the original image is followed by 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.

The variants of HOG exhibit variable cell shapes, and it has been observed that one way to accelerate the computation of the histogram of phase magnitudes is to leverage the integral image once again. Therefore, straddling the line between HOG and Haar features, the Integral Channel Features have recently demonstrated interesting performance, effectively serving as a generalization of HOG that utilizes the integral image.

The characteristic values that can be extracted derive from the summation of areas calculated not directly on the original image but from different secondary images, obtained through non-linear processing of the area to be characterized. Among the possible processing techniques, the most common are the gradient phase channels already seen in HOG, the gradient magnitude image, the grayscale image itself, and, if available, two additional channels representing chrominance.

Regarding the gradient, it is often calculated using the Sobel operator, but various experiments show that a simple derivative filter produces satisfactory results nonetheless. In this case, the Sobel phase can be used with or without sign, depending on the specific applications.

The scalar representing the characteristic to be extracted is simply the sum of a rectangular area within one of the calculated channels.