It used to be able to output uint8 histogram, whose max pixel counts of 255 could easily overflow. Addressed by limiting the output type to float.
Normalized histograms are now generated, otherwise behaviour is unpredictable at boundaries or when pixels are not permitted by a mask.
The new optional n_bins parameter allows the caller to specify the size of the histogram generated. Having it fixed to image.max()+1 could result in feature vectors being shorter than desired just due to a value not being used in an image.
An example has been added to the docs, that demonstrates the application of windows histograms in object matching; a single coin is extracted and found by chi squared histogram matching.
Geoffrey French