Fix #1452: use "true" original for HDR metrics

Thanks for fix. Any HDR result on the impact on the efficiency?

As far as I understand HDRtools can be used either directly inside VTM or as a standalone app. Currently, results do not match between the two ways of using it. I don't know how big the difference is, but it doesn't really matter because there should be no difference.

I have not checked 360Lib, but sounds like a problems that could occur there as well?

For 360Lib, the issue might be in the library itself. The VTM has:

#if EXTENSION_360_VIDEO
  m_ext360.calculatePSNRs(pcPic);
#endif

If I recall correctly, the difference occurs due to single precision being used in HDRTools (standalone) and double precision (inside VTM).

Minor differences in performance can be observed between the Anchor measurements and the new VTM integrated measurements. These relate to the distortion introduced by the conversion to the half float EXR representation that is utilized by the Anchor data, which is avoided by the new method. In a sense, the new method is also more accurate.

From JVET-O0756

@dmehlem Thanks for pointing that out. However, the discrepancy being addressed here is much larger than a rounding error.

approved this merge request

merged

mentioned in commit 88ec22ab