Support HDR DNG images from HDRMerge

[Beep6581]

I will paste the correspondence I had via email, as it is good to have this public so that other people can take part in the discussion.

On 8/14/2014 1:32 PM, Beep6581 wrote:
Hi

HDRMerge is an awesome FOSS program from Javier Celaya for generating a high
dynamic range raw image in the DNG format from a bracketed raw set.
https://github.com/jcelaya/hdrmerge

Please add support for correctly reading the thumbnails and embedded preview
from these raw images. They follow standards and should therefore be
supported.
You can download a sample image here:
http://rawtherapee.com/shared/test_images/hdrmerge_045.dng

I was informed I should ask you directly:
https://bugs.kde.org/show_bug.cgi?id=338081
As I can't find your bug tracker anywhere, I'm doing this via direct contact.
If you have one, please file the request on my behalf.

Kind regards

On 2014-08-14 04:57, Alex Tutubalin wrote:

Hi,

LibRaw's main goal is RAW data decoding, so LibRaw::open_file() call
will return 'success' only if known RAW data is found in file.
Unfortunately, high-bit integer and floating point RAWs (DNGs) are not
supported due to LibRaw's internals specifics.

So, there is no way to support thumbnails/previews in unsupported
format without API break.

You may use some EXIF management library (e.g. exiv2) to extract
metadata and previews.

Alex Tutubalin, LibRaw team

These HDR DNG images are raw, they are not demosaiced. Furthermore, they are structured "by the book" as per the DNG specification. This includes the embedded JPEG thumbnail and preview.
My request is that libraw at least allows the extraction of these embedded JPEG thumbnails and preview images (I was lead to believe by Caulier Gilles that digiKam relies on libraw to extract these embedded JPEGs - is this correct?) if raw decoding and demosaicing is not possible.

Kind regards

[LibRaw]

Yes, these DNGs are 'by the book' (at least, exiftool dumps the metadata).
But these DNGs are 16-bit floating point, this format is not supported by LibRaw. And no way to support it easy because it requires heavy internals modifications.

If someone will contribute patch to support FP RAWs in LibRaw, I'll be happy.

[Beep6581]

But I'm not requesting support for decoding the raw data block it contains, just for the embedded JPEG thumbnail and preview. That makes the high bit depth floating point issue irrelevant as far as this request is concerned.

[alextutubalin]

'Supporting' some RAW format (e.g. floating point DNGs) without real support will break existing programs.

[Beep6581]

Hi again, I wanted to point this out:
"(...) They fail, because libraw does not know how to read the float data. Since reading the data and the preview is performed in just one call to open(), the preview is not available when reading the data fails."
jcelaya/hdrmerge#27 (comment)

Would be nice if libraw could read the embedded preview or thumb even if it cannot read the raw data.

[LibRaw]

Again: 'Supporting' some RAW format (e.g. floating point DNGs) without real support will break existing programs.

[DavidOliver]

I understand proper floating point support is perhaps not viable, and I'm just registering my coming up against the limitation. I've started to try out HDRMerge and I use Darktable for processing, which can't import the resultant HDR files:

failed to read camera white balance information from (filename)

Anyway, thanks for libraw!

[LibRaw]

Darktable do not use LibRaw anymore, so FP-DNG support in LibRaw will not help you.

Current LibRaw supports FP-DNG if compiled with Adobe DNG SDK.

[Beep6581]

It is good to hear that LibRaw can now support floating-point DNGs, but unless I missed something, there is no mention of this anywhere in any of the docs. It would be good necessary to mention this along with instructions how to compile LibRaw with Adobe DNG SDK, at least for the sake of being able to point people to that paragraph.

[LibRaw]

Quote from Changelog:

• LibRaw can be built with Adobe DNG SDK support (you'll need to compile
  this SDK yourself for your system).
  To compile/build with DNG SDK specify USE_DNGSDK in compiler flags (you may also need to adjust include path and linker settings).
  On Runtime:
  • create dng_host (or derived class) entity in your program;
  • pass it to LibRaw:set_dng_host(dng_host *) call to enable DNG SDK use on runtime
  • adjust imgdata.params.use_dngsdk parameter with one of this values
    (ORed bitwise):
    LIBRAW_DNG_NONE - do not use DNG SDK for DNG processing
    LIBRAW_DNG_FLOAT - process floating point DNGs via Adobe DNG SDK
    LIBRAW_DNG_LINEAR - process 'linear DNG' (i.e. demosaiced)
    LIBRAW_DNG_DEFLATE - process DNGs compressed with deflate (gzip)
    LIBRAW_DNG_XTRANS - process Fuji X-Trans DNGs (6x6 CFA)
    LIBRAW_DNG_OTHER - process other files (so usual 2x2 bayer)
    Default value for this field:
    LIBRAW_DNG_DEFAULT=LIBRAW_DNG_FLOAT|LIBRAW_DNG_LINEAR|LIBRAW_DNG_DEFLATE
    You also may use this macro to pass all supported DNGs to Adobe SDK:
    LIBRAW_DNG_ALL = LIBRAW_DNG_FLOAT|LIBRAW_DNG_LINEAR|LIBRAW_DNG_XTRANS|LIBRAW_DNG_OTHER
• New API calls
  unsigned LibRaw::capabilities and C-API libraw_capabilities()
  allows developers to determine LibRaw compile flags at runtime.
  Returns ORed bit fields:
  LIBRAW_CAPS_RAWSPEED - LibRaw was compiled with RawSpeed Support
  LIBRAW_CAPS_DNGSDK - LibRaw was compiled with Adobe DNG SDK
  LIBRAW_CAPS_DEMOSAICSGPL2, LIBRAW_CAPS_DEMOSAICSGPL3 - LibRaw was compiled with demosaic packs (GPL2/GPL3)
• floating point support:
  • new data fields:
    imgdata.rawdata.float_image - bayer float data
    imgdata.rawdata.float3_image - 3-component float data
    imgdata.rawdata.float4_image - 4-component float data
    imgdata.color.fmaximum - float data maximum (calculated from real data, rounded to 1.0 if below 1.0)
  • new raw processing flag
    LIBRAW_PROCESSING_CONVERTFLOAT_TO_INT - converts float data to 16-bit integer immediately after decoding with default parameters
  • new API Calls:
    int LibRaw::is_floating_point() returns non-zero if RAW file contains floating point data
    int LibRaw::have_fpdata() returns non-zero if rawdata.float*_image is not null (so FP data has unpacked but not converted to integrer, see below).
    LibRaw::convertFloatToInt(float dmin=4096.f, float dmax=32767.f, float dtarget = 16383.f) converts float/float3/float4_image to raw_image/color3/color4_image with or withou scaling:
    • if both real data maximum and metadata maximum are within range ( >= dmin && <=dmax), float data are just converted to int
    • if data is out of range noted above, values are scaled so real data maximum becomes dtarget
    • if data was rescaled (normalized), scale multiplier is stored in imgdata.color.fnorm

[Beep6581]

I would suggest putting that in README or INSTALL.

[LibRaw]

Yes, when current version becomes -stable, this text to be propagated to INSTALL, ./configure and constant description

[LebedevRI]

@DavidOliver as for darktable, i will be looking into rawspeed support for hdrmerge-produced DNG's in the coming ${timeframe}.
Sad thing: we were pretty sure hdrmerge could produce uncompressed DNG's, but it does not. I do have patch, but apparently it too needs changes in rawspeed.