Re: [Vxl-maintainers] vil_convert.h RGB to grayscale conversion numerical issue

[Ian S. <ian...@st...>]

Wheeler, Frederick W (Research) wrote:
> VXL Maintainers:
> 
> Here is a recent change to ITK by Bill Lorensen for the conversion of RGB to
> grayscale:
> 
> http://www.itk.org/cgi-bin/viewcvs.cgi/Code/IO/itkConvertPixelBuffer.txx?roo
> t=Insight&r1=1.15&r2=1.16
> 
> Here is a snippet:
> 
> 139,141c141,143
> <       0.2125 * static_cast<OutputComponentType>(*inputData) +
> <       0.7154 * static_cast<OutputComponentType>(*(inputData+1)) +
> <       0.0721 * static_cast<OutputComponentType>(*(inputData+2))   );
> ---
> 
>>      (2125.0 * static_cast<OutputComponentType>(*inputData) +
>>       7154.0 * static_cast<OutputComponentType>(*(inputData+1)) +
>>       0721.0 * static_cast<OutputComponentType>(*(inputData+2))) /
> 
> 10000.0   );
> 
> It was found that this change was necessary to get the same results on Intel
> and non-Intel platforms when r=g=b.  With q=0.2125*r+0.7154*g+0.0721*b and
> r=g=b=187 the result is q==186 on non-Intel platforms.  There are a few
> other values from 0-255 that give similar results.
> 
> In vil_convert.h we cannot make the same change without adding computation
> because vil_convert_rgb_to_grey lets you pass in the weights.  We'd have to
> multiply by 10000 to maintain the current API.
> 
> Should VXL make the same change?  I don't think so because of the additional
> computation, but I wanted to report the issue.

I think, it might be worth pointing this out in the documentation, but 
leaving the code as is.

Ian.