[libcamera-devel] [PATCH v2 3/3] libcamera: qcam: Improve colour information in DNG files

[Laurent Pinchart]

Hi David,

Thank you for the patch.

On Fri, Jul 24, 2020 at 03:56:18PM +0100, David Plowman wrote:
> This patch improves the colour information recorded in DNG files using
> the ColourCorrectionMatrix metadata for the image. Note that we are
> not supplying a full calibration using two illuminants, nonetheless
> the single matrix here appears to be respected by a number of tools.
> 
> Signed-off-by: David Plowman <david.plowman at raspberrypi.com>
> ---
>  src/qcam/dng_writer.cpp | 138 ++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 138 insertions(+)
> 
> diff --git a/src/qcam/dng_writer.cpp b/src/qcam/dng_writer.cpp
> index 61505d3..4f638ec 100644
> --- a/src/qcam/dng_writer.cpp
> +++ b/src/qcam/dng_writer.cpp
> @@ -34,6 +34,97 @@ struct FormatInfo {
>  			      unsigned int stride);
>  };
>  
> +struct Matrix3d {
> +	Matrix3d()
> +	{
> +	}
> +
> +	Matrix3d(float m0, float m1, float m2,
> +		 float m3, float m4, float m5,
> +		 float m6, float m7, float m8)
> +	{
> +		m[0] = m0, m[1] = m1, m[2] = m2;
> +		m[3] = m3, m[4] = m4, m[5] = m5;
> +		m[6] = m6, m[7] = m7, m[8] = m8;
> +	}
> +
> +	Matrix3d(Span<const float> const &span)

We usually put the const keyword before the type.

> +		: Matrix3d(span[0], span[1], span[2],
> +			   span[3], span[4], span[5],
> +			   span[6], span[7], span[8])
> +	{
> +	}
> +
> +	static Matrix3d diag(float diag0, float diag1, float diag2)
> +	{
> +		return Matrix3d(diag0, 0, 0, 0, diag1, 0, 0, 0, diag2);
> +	}
> +
> +	static Matrix3d identity()
> +	{
> +		return Matrix3d(1, 0, 0, 0, 1, 0, 0, 0, 1);
> +	}
> +
> +	Matrix3d transpose() const
> +	{
> +		return { m[0], m[3], m[6], m[1], m[4], m[7], m[2], m[5], m[8] };
> +	}
> +
> +	Matrix3d cofactors() const
> +	{
> +		return { m[4] * m[8] - m[5] * m[7],
> +			 -(m[3] * m[8] - m[5] * m[6]),
> +			 m[3] * m[7] - m[4] * m[6],
> +			 -(m[1] * m[8] - m[2] * m[7]),
> +			 m[0] * m[8] - m[2] * m[6],
> +			 -(m[0] * m[7] - m[1] * m[6]),
> +			 m[1] * m[5] - m[2] * m[4],
> +			 -(m[0] * m[5] - m[2] * m[3]),
> +			 m[0] * m[4] - m[1] * m[3] };
> +	}
> +
> +	Matrix3d adjugate() const
> +	{
> +		return cofactors().transpose();
> +	}
> +
> +	float determinant() const
> +	{
> +		return m[0] * (m[4] * m[8] - m[5] * m[7]) -
> +		       m[1] * (m[3] * m[8] - m[5] * m[6]) +
> +		       m[2] * (m[3] * m[7] - m[4] * m[6]);
> +	}
> +
> +	Matrix3d inverse() const
> +	{
> +		return adjugate() * (1.0 / determinant());
> +	}
> +
> +	Matrix3d operator*(Matrix3d const &other) const

Same here.

> +	{
> +		Matrix3d result;
> +		for (unsigned int i = 0; i < 3; i++) {
> +			for (unsigned int j = 0; j < 3; j++) {
> +				result.m[i * 3 + j] =
> +					m[i * 3 + 0] * other.m[0 + j] +
> +					m[i * 3 + 1] * other.m[3 + j] +
> +					m[i * 3 + 2] * other.m[6 + j];
> +			}
> +		}
> +		return result;
> +	}
> +
> +	Matrix3d operator*(float f) const
> +	{
> +		Matrix3d result;
> +		for (unsigned int i = 0; i < 9; i++)
> +			result.m[i] = m[i] * f;
> +		return result;
> +	}
> +
> +	float m[9];
> +};

A couple of comments if we later want a 3x3 matrix implementation in
libcamera:

- The inverse function should guard against division by 0
- Most functions should be constexpr

Not something we need to address now.

> +
>  void packScanlineSBGGR10P(void *output, const void *input, unsigned int width)
>  {
>  	const uint8_t *in = static_cast<const uint8_t *>(input);
> @@ -315,6 +406,53 @@ int DNGWriter::write(const char *filename, const Camera *camera,
>  	TIFFSetField(tif, TIFFTAG_PLANARCONFIG, PLANARCONFIG_CONTIG);
>  	TIFFSetField(tif, TIFFTAG_SAMPLEFORMAT, SAMPLEFORMAT_UINT);
>  
> +	/*
> +	 * Fill in some reasonable colour information in the DNG. We supply
> +	 * the "neutral" colour values which determine the white balance, and the
> +	 * "ColorMatrix1" which converts XYZ to (un-white-balanced) camera RGB.
> +	 * Note that this is not a "proper" colour calibration for the DNG,
> +	 * nonetheless, many tools should be able to render the colours better.
> +	 */
> +	float neutral[3] = { 1, 1, 1 };
> +	Matrix3d wbGain = Matrix3d::identity();
> +	/* From http://www.brucelindbloom.com/index.html?Eqn_RGB_XYZ_Matrix.html */
> +	const Matrix3d rgb2xyz(0.4124564, 0.3575761, 0.1804375,
> +			       0.2126729, 0.7151522, 0.0721750,
> +			       0.0193339, 0.1191920, 0.9503041);
> +	Matrix3d ccm = Matrix3d::identity();
> +	/*
> +	 * Pick a reasonable number eps to protect against singularities. It
> +	 * should be comfortably larger than the point at which we run into
> +	 * numerical trouble, yet smaller than any plausible gain that we might
> +	 * apply to a colour, either explicitly or as part of the colour matrix.
> +	 */
> +	const double eps = 1e-2;
> +
> +	if (metadata.contains(controls::ColourGains)) {
> +		Span<const float> const &colourGains = metadata.get(controls::ColourGains);
> +		if (colourGains[0] > eps && colourGains[1] > eps) {
> +			wbGain = Matrix3d::diag(colourGains[0], 1, colourGains[1]);
> +			neutral[0] = 1.0 / colourGains[0]; /* red */
> +			neutral[2] = 1.0 / colourGains[1]; /* blue */
> +		}
> +	}
> +	if (metadata.contains(controls::ColourCorrectionMatrix)) {
> +		Span<const float> const &coeffs = metadata.get(controls::ColourCorrectionMatrix);
> +		Matrix3d ccm_supplied(coeffs);

s/ccm_supplied/ccmSupplied/

Reviewed-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>

I've addressed those small issues when applying the patches, I'll push
right after testing.

> +		if (ccm_supplied.determinant() > eps)
> +			ccm = ccm_supplied;
> +	}
> +
> +	/*
> +	 * rgb2xyz is known to be invertible, and we've ensured above that both
> +	 * the ccm and wbGain matrices are non-singular, so the product of all
> +	 * three is guaranteed to be invertible too.
> +	 */
> +	Matrix3d colorMatrix1 = (rgb2xyz * ccm * wbGain).inverse();
> +
> +	TIFFSetField(tif, TIFFTAG_COLORMATRIX1, 9, colorMatrix1.m);
> +	TIFFSetField(tif, TIFFTAG_ASSHOTNEUTRAL, 3, neutral);
> +
>  	/*
>  	 * Reserve space for the SubIFD and ExifIFD tags, pointing to the IFD
>  	 * for the raw image and EXIF data respectively. The real offsets will

-- 
Regards,

Laurent Pinchart