I was reading OpenCV documentation about histogram backprojection where I came across this line:
Also, the object histogram should be normalized before passing on to the backproject function.
I am interested as to why the histogram should be normalized? Does it make the algorithm function better, as if so, why is that?
I tried understanding the documentation on my own, plotted the histograms, saw the formula for normalization, but I didn't see benefits from it. Any help would be appreciated, thanks.
It is not only this application where normalisation is desired. It is done in many other applications as well. Normalisation scales the histogram so that its values sum to a specific total (e.g., 1, 255, or some other constant). Here’s why this matters:
1. Consistency Across Different Region of Interests (ROIs)
The raw histogram’s bin values depend on the number of pixels in the ROI. A larger ROI will have higher bin counts than a smaller ROI, even if the color distribution is identical. Without normalization, the backprojection output would scale with the ROI size, making it hard to compare results across different objects or images. Normalizing ensures the histogram represents a probability distribution (relative frequencies), independent of ROI size.
2. Scaling for Algorithm Output
cv2.calcBackProject produces an 8-bit single-channel image (values 0–255). The histogram bin values are directly mapped to pixel intensities in this output. If the histogram isn’t normalized, bin values could exceed 255 (for large ROIs) or be very small (for tiny ROIs), leading to clipping or faint outputs. Normalizing to a range like [0, 255] ensures the backprojection image uses the full dynamic range effectively, making the result visually meaningful and consistent.