The amount of color change is about -(2,2,2)
My tests:
Take an AVC1 coded movie. Transcode to FFV1 lossless (e.g. by Avidemux).
VideoCapture a movie and VideoWrite it again, with all 4 combinations:
FFV1 ==> AVC1, AVC1 ==> FFV1, etc.
Then export the first frame of the pairs respectively (by ffmpeg CLI):
ffmpeg -i input -frames:v 1 ffmpeg_screenshot.png
and compare the frames by the averages of their BGR and HSV channels:
(Stripped without slip I hope)
Mat imageMat, reReadMat,
subMat, imageMatHSV, reReadMatHSV;
imageMat = imread ("ScreenShotA.png");
reReadMat = imread ("ScreenShotB.png");
cout << endl << "Averages:" << endl;
cv::Scalar meanVal = mean (imageMat);
cout << "A:\t Orig: B = " << meanVal.val[0] << "\tG = " << meanVal.val[1] << "\tR = " << meanVal.val[2] << endl;
meanVal = mean (reReadMat);
cout << "B:\t Orig: B = " << meanVal.val[0] << "\tG = " << meanVal.val[1] << "\tR = " << meanVal.val[2] << endl << endl;
// Now the both differences:
subMat = imageMat - reReadMat;
meanVal = mean (subMat);
cout << "A-B:\t Orig: B = " << meanVal.val[0] << "\tG = " << meanVal.val[1] << "\tR = " << meanVal.val[2] << endl;
subMat = reReadMat - imageMat;
cv::Scalar meanValBminusA = mean (subMat);
cout << "B-A:\t Orig: B = " << meanValBminusA.val[0] << "\tG = " << meanValBminusA.val[1] << "\tR = " << meanValBminusA.val[2] << endl << endl;
// mean (A-B) - mean (B-A), "difference of differences" (DOD)
meanVal -= meanValBminusA;
cout << "DOD:\t Orig: B = " << meanVal.val[0] << "\tG = " << meanVal.val[1] << "\tR = " << meanVal.val[2] << endl << endl << endl;
// Now HSV channels
cvtColor (imageMat, imageMatHSV, COLOR_BGR2HSV);
cvtColor (reReadMat, reReadMatHSV, COLOR_BGR2HSV);
meanVal = mean (imageMatHSV);
cout << "A:\t Orig: H = " << meanVal.val[0] << "\tS = " << meanVal.val[1] << "\tV = " << meanVal.val[2] << endl;
meanVal = mean (reReadMatHSV);
cout << "B:\t Orig: H = " << meanVal.val[0] << "\tS = " << meanVal.val[1] << "\tV = " << meanVal.val[2] << endl << endl;
// Now the both differences:
subMat = imageMatHSV - reReadMatHSV;
meanVal = mean (subMat);
cout << "A-B:\t Orig: H = " << meanVal.val[0] << "\tS = " << meanVal.val[1] << "\tV = " << meanVal.val[2] << endl;
subMat = reReadMatHSV - imageMatHSV;
meanValBminusA = mean (subMat);
cout << "B-A:\t Orig: H = " << meanValBminusA.val[0] << "\tS = " << meanValBminusA.val[1] << "\tV = " << meanValBminusA.val[2] << endl << endl;
// mean (A-B) - mean (B-A), "difference of differences"
meanVal -= meanValBminusA;
cout << "DOD:\t Orig: H = " << meanVal.val[0] << "\tS = " << meanVal.val[1] << "\tV = " << meanVal.val[2] << endl << endl;
The outputs
FFV1 ==> AVC1 (VideoWriter impact):
A: Orig: B = 60.2154 G = 62.571 R = 71.093
B: Orig: B = 57.9042 G = 61.0538 R = 69.0174
A-B: Orig: B = 2.53997 G = 1.61761 R = 2.15573
B-A: Orig: B = 0.228726 G = 0.100314 R = 0.0800911
DOD: Orig: B = 2.31124 G = 1.51729 R = 2.07564
A: Orig: H = 76.933 S = 69.9563 V = 72.1645
B: Orig: H = 71.4863 S = 75.3879 V = 70.1948
A-B: Orig: H = 7.79836 S = 1.45931 V = 2.05884
B-A: Orig: H = 2.35164 S = 6.89095 V = 0.0891775
DOD: Orig: H = 5.44672 S = -5.43164 V = 1.96966
AVC1 ==> AVC1 (VideoCapture plus VideoWriter impact)
is the same as above (FFV1 ==> AVC1).
FFV1 ==> FFV1
no changes, identical image content.
(BTW: imwrite does the same image content).
AVC1 ==> FFV1 (VideoCapture impact):
no changes, identical image content.
(BTW: imwrite does the same image content).
Checking visually the pixel-wise difference of the pairs above
reveals that it’s always the RGB channels that are about homogenous.
The HSV channels’ differences are strongly structured.
Conclusion:
VideoCapture does fine for AVC1 (and FFV1).
VideoWrite for AVC1 movie shall be done as:
if (bAVC1Write)
outputVideo.write (frame + VideoWriterCorretion4AVC1);
else
outputVideo.write (frame);
with
const Scalar VideoWriterCorretion4AVC1 (2,2,2);
With that correction I did the test FFV1 ==> AVC1 again:
A: Orig: B = 60.2154 G = 62.571 R = 71.093
B: Orig: B = 59.8431 G = 62.9274 R = 70.9557
A-B: Orig: B = 0.892717 G = 0.314124 R = 0.566224
B-A: Orig: B = 0.520381 G = 0.670454 R = 0.428946
DOD: Orig: B = 0.372336 G = -0.356329 R = 0.137278
A: Orig: H = 76.933 S = 69.9563 V = 72.1645
B: Orig: H = 71.7396 S = 70.4345 V = 72.13
A-B: Orig: H = 7.52111 S = 2.98839 V = 0.506662
B-A: Orig: H = 2.32764 S = 3.46663 V = 0.472235
DOD: Orig: H = 5.19347 S = -0.47824 V = 0.0344271
That’s better I might say.
However it’s not the perfect correction, just best overall.
There are colors that need different correction.
I’m well aware that a lossy codec (decoded, encoded) like AVC1
will produce differences.
But they won’t be as systematically and as large as this.
So there’s a bug awaiting for unveiling.
For scientific work lossless codecs will be used if possible;
but if there are big movies you’ll have to use compression.
I haven’t tested other codecs.
In my OpenCV 4.8.0 I used a newer ffmpeg (5.0.1).
I can’t tell about the (prehistoric) default one.
Confirmation/ contradiction on that?
EDIT:
mean() of Hue channel cannot be calculated as arithmetical mean, as Hue’s values are circular. Must be done as here.
But H channel aren’t of any importance in this case, due to structural visual difference (see above).