Messages

Yes, totally right. A denoising algo would be best. I could test this algo in MLVApp, because it was easy to adapt to what we need:
http://www.ipol.im/pub/art/2011/bcm_nlm/
It works but it is soooo slow - OMG - some minutes per frame, and only on Win & Linux multithreaded to work correctly. So no real solution for us

I've found this implementation of BM3D that according to the author "can be used in real-time video denoising", using CUDA:
https://github.com/JeffOwOSun/gpu-bm3d

There's also this page linking to many denoise code and papers:
https://github.com/wenbihan/reproducible-image-denoising-state-of-the-art

Just curious: does MLVApp autodetects white levels or use predefined values?
@a1ex explained why this would be a good idea:
https://www.magiclantern.fm/forum/index.php?topic=10111.msg98298#msg98298
https://www.magiclantern.fm/forum/index.php?topic=10111.msg98359#msg98359

You need a installed Qt and the source code. You load the .pro into QtCreator, setup the debug mode, press debug, do the crash, and Qt tells you where it crashes. For me it also does not crash on Windows...

Ok, I'm trying.

Found it! They call it "Bayer drizzle". The implementation seems to be from Dave Coffin (from dcraw). The software DeepSkyStacker uses it.
Would be nice to have it on HDRMerge. I'll see if Beep6581 have more information on this.

Nevermind about that. Guess who's just been robbed of all his gear? Yeah. Next camera I'll get is a blackmagic pocket 4k.
Loved to work with the 5d and magic lantern. Very rewarding and fun. Made things feel special.

Holy shit. I'm sorry.
Be careful with BPCC. The sensor size is really small. Maybe get a metabones adapter. It is also heavier, so check your stabilization equipment (steadycams) if they can handle it.

Take care of your belongings guys and have insurance.

Yeah. If I was robbed it would be game over for me. I wouldn't have the money to buy everything again.

@Danne @garry23 Thanks for sharing. But, I think these techniques presented just use a variation of noise-averaging. My idea was to use this pre-demosaicing, to get full color information...

as far i can understand these fragments, the idea is to
- shoot [n] images
- detect motion between them (optical flow)
- overlay the raw images accordingly

if you have some motion now, due to vibration etc, different bayer pixels overlay now and you have a good estimate what R, G and B is at a specific "image" location.

Yes, that's the idea...

wonder if it was meant to shift the sensor automatically (didnt hasselblad do this already using piezo?) or if the unsteady hands would deliver the required motion.
latter makes sense if you have motion *estimation* because with piezo you already know the motion.
we could capture "short 4-frame MLV files" like with silent picture and the post can be done on the pc, but...

Manual motion.

*AFTER* someone has proven that it makes sense

If you have any research on this, hit me please
I had this idea after having issues with demosaicing from the images generated on HDRMerge. So I thought that, maybe, debayering would not be necessary if you have enough images with a shift between them...

@masc @bouncyball How can I debug? I'm using the binary from github. Tried to execute from terminal, but it's quiet (no verbose info, even with --debug). The binary seems to use QT 5.9.1. I don't know how to compile on windows, just on linux, and I don't think this issue happens on linux (from what you're saying)...

Does anyone know a software or research paper implementing something like this?

Last release for Windows is crashing during batch export (ProRes 444), while using darkframe subtraction
Darkframe sample (averaged using MLVApp):
https://minfil.com/S8b8tdf3b6/50d_iso_100.MLV

ps: it renders ~3 files on the batch, then crashs.

If it blurs the image too much, it wouldn't be of much use.  It should preserve fine detail as it does in DaVinci Resolve for example.  Also, it should allow a very fine adjustment as this is the case with the blur radius slider.  This will make it possible to clean the most intrusive monochrome noise while leaving the rest in the image for a more cinematic look.  I will be looking forward to it!  Thanks, Masc.

A proper denoising algo would be better, no? This and this implementations are GPL licensed, maybe it could be adapted. Just an idea.

I found an implementation of LMMSE debayer with correct interface in C, so I added it to MLVApp now. Has someone used it in past in any other program?

I have on RawTherapee...

I have the problem that hard contrast lines are getting very green and cyan. Is that normal? Only highiso clips are okay... but AMaZE looks nearly always way better.

Yep. AMaZE is always better than LMMSE. At least on RT.
The only demosaicing algorithm on RT that is better than AMaZE is IGV (Integrated Gaussian Vector), but only for highly noisy images. This is the code on RT:

Code Select Expand


/***
*
*   Bayer CFA Demosaicing using Integrated Gaussian Vector on Color Differences
*   Revision 1.0 - 2013/02/28
*
*   Copyright (c) 2007-2013 Luis Sanz Rodriguez
*   Using High Order Interpolation technique by Jim S, Jimmy Li, and Sharmil Randhawa
*
*   Contact info: [email protected]
*
*   This code is distributed under a GNU General Public License, version 3.
*   Visit <http://www.gnu.org/licenses/> for more information.
*
***/
// Adapted to RawTherapee by Jacques Desmis 3/2013
// SSE version by Ingo Weyrich 5/2013
#ifdef __SSE2__
#define CLIPV(a) LIMV(a,zerov,c65535v)
void RawImageSource::igv_interpolate(int winw, int winh)
{
    static const float eps = 1e-5f, epssq = 1e-5f; //mod epssq -10f =>-5f Jacques 3/2013 to prevent artifact (divide by zero)

    static const int h1 = 1, h2 = 2, h3 = 3, h5 = 5;
    const int width = winw, height = winh;
    const int v1 = 1 * width, v2 = 2 * width, v3 = 3 * width, v5 = 5 * width;
    float* rgb[2];
    float* chr[4];
    float *rgbarray, *vdif, *hdif, *chrarray;
    rgbarray    = (float (*)) malloc((width * height) * sizeof( float ) );
    rgb[0] = rgbarray;
    rgb[1] = rgbarray + (width * height) / 2;

    vdif  = (float (*)) calloc( width * height / 2, sizeof * vdif );
    hdif  = (float (*)) calloc( width * height / 2, sizeof * hdif );

    chrarray    = (float (*)) calloc( width * height, sizeof( float ) );
    chr[0] = chrarray;
    chr[1] = chrarray + (width * height) / 2;

    // mapped chr[2] and chr[3] to hdif and hdif, because these are out of use, when chr[2] and chr[3] are used
    chr[2] = hdif;
    chr[3] = vdif;

    border_interpolate2(winw, winh, 7, rawData, red, green, blue);

    if (plistener) {
        plistener->setProgressStr (Glib::ustring::compose(M("TP_RAW_DMETHOD_PROGRESSBAR"), RAWParams::BayerSensor::getMethodString(RAWParams::BayerSensor::Method::IGV)));
        plistener->setProgress (0.0);
    }

#ifdef _OPENMP
    #pragma omp parallel default(none) shared(rgb,vdif,hdif,chr)
#endif
    {
        __m128 ngv, egv, wgv, sgv, nvv, evv, wvv, svv, nwgv, negv, swgv, segv, nwvv, nevv, swvv, sevv, tempv, temp1v, temp2v, temp3v, temp4v, temp5v, temp6v, temp7v, temp8v;
        __m128 epsv = _mm_set1_ps( eps );
        __m128 epssqv = _mm_set1_ps( epssq );
        __m128 c65535v = _mm_set1_ps( 65535.f );
        __m128 c23v = _mm_set1_ps( 23.f );
        __m128 c40v = _mm_set1_ps( 40.f );
        __m128 c51v = _mm_set1_ps( 51.f );
        __m128 c32v = _mm_set1_ps( 32.f );
        __m128 c8v = _mm_set1_ps( 8.f );
        __m128 c7v = _mm_set1_ps( 7.f );
        __m128 c6v = _mm_set1_ps( 6.f );
        __m128 c10v = _mm_set1_ps( 10.f );
        __m128 c21v = _mm_set1_ps( 21.f );
        __m128 c78v = _mm_set1_ps( 78.f );
        __m128 c69v = _mm_set1_ps( 69.f );
        __m128 c3145680v = _mm_set1_ps( 3145680.f );
        __m128 onev = _mm_set1_ps ( 1.f );
        __m128 zerov = _mm_set1_ps ( 0.f );
        __m128 d725v = _mm_set1_ps ( 0.725f );
        __m128 d1375v = _mm_set1_ps ( 0.1375f );

        float *dest1, *dest2;
        float ng, eg, wg, sg, nv, ev, wv, sv, nwg, neg, swg, seg, nwv, nev, swv, sev;
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 0; row < height - 0; row++) {
            dest1 = rgb[FC(row, 0) & 1];
            dest2 = rgb[FC(row, 1) & 1];
            int col, indx;

            for (col = 0, indx = row * width + col; col < width - 7; col += 8, indx += 8) {
                temp1v = LVFU( rawData[row][col] );
                temp1v = CLIPV( temp1v );
                temp2v = LVFU( rawData[row][col + 4] );
                temp2v = CLIPV( temp2v );
                tempv = _mm_shuffle_ps( temp1v, temp2v, _MM_SHUFFLE( 2, 0, 2, 0 ) );
                _mm_storeu_ps( &dest1[indx >> 1], tempv );
                tempv = _mm_shuffle_ps( temp1v, temp2v, _MM_SHUFFLE( 3, 1, 3, 1 ) );
                _mm_storeu_ps( &dest2[indx >> 1], tempv );
            }

            for (; col < width; col++, indx += 2) {
                dest1[indx >> 1] = CLIP(rawData[row][col]); //rawData = RT datas
                col++;
                if(col < width)
                    dest2[indx >> 1] = CLIP(rawData[row][col]); //rawData = RT datas
            }
        }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.13);
            }
        }

#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 5; row < height - 5; row++) {
            int col, indx, indx1;

            for (col = 5 + (FC(row, 1) & 1), indx = row * width + col, indx1 = indx >> 1; col < width - 12; col += 8, indx += 8, indx1 += 4) {
                //N,E,W,S Gradients
                ngv = (epsv + (vabsf(LVFU(rgb[1][(indx - v1) >> 1]) - LVFU(rgb[1][(indx - v3) >> 1])) + vabsf(LVFU(rgb[0][indx1]) - LVFU(rgb[0][(indx1 - v1)]))) / c65535v);
                egv = (epsv + (vabsf(LVFU(rgb[1][(indx + h1) >> 1]) - LVFU(rgb[1][(indx + h3) >> 1])) + vabsf(LVFU(rgb[0][indx1]) - LVFU(rgb[0][(indx1 + h1)]))) / c65535v);
                wgv = (epsv + (vabsf(LVFU(rgb[1][(indx - h1) >> 1]) - LVFU(rgb[1][(indx - h3) >> 1])) + vabsf(LVFU(rgb[0][indx1]) - LVFU(rgb[0][(indx1 - h1)]))) / c65535v);
                sgv = (epsv + (vabsf(LVFU(rgb[1][(indx + v1) >> 1]) - LVFU(rgb[1][(indx + v3) >> 1])) + vabsf(LVFU(rgb[0][indx1]) - LVFU(rgb[0][(indx1 + v1)]))) / c65535v);
                //N,E,W,S High Order Interpolation (Li & Randhawa)
                //N,E,W,S Hamilton Adams Interpolation
                // (48.f * 65535.f) = 3145680.f
                tempv = c40v * LVFU(rgb[0][indx1]);
                nvv = LIMV(((c23v * LVFU(rgb[1][(indx - v1) >> 1]) + c23v * LVFU(rgb[1][(indx - v3) >> 1]) + LVFU(rgb[1][(indx - v5) >> 1]) + LVFU(rgb[1][(indx + v1) >> 1]) + tempv - c32v * LVFU(rgb[0][(indx1 - v1)]) - c8v * LVFU(rgb[0][(indx1 - v2)]))) / c3145680v, zerov, onev);
                evv = LIMV(((c23v * LVFU(rgb[1][(indx + h1) >> 1]) + c23v * LVFU(rgb[1][(indx + h3) >> 1]) + LVFU(rgb[1][(indx + h5) >> 1]) + LVFU(rgb[1][(indx - h1) >> 1]) + tempv - c32v * LVFU(rgb[0][(indx1 + h1)]) - c8v * LVFU(rgb[0][(indx1 + h2)]))) / c3145680v, zerov, onev);
                wvv = LIMV(((c23v * LVFU(rgb[1][(indx - h1) >> 1]) + c23v * LVFU(rgb[1][(indx - h3) >> 1]) + LVFU(rgb[1][(indx - h5) >> 1]) + LVFU(rgb[1][(indx + h1) >> 1]) + tempv - c32v * LVFU(rgb[0][(indx1 - h1)]) - c8v * LVFU(rgb[0][(indx1 - h2)]))) / c3145680v, zerov, onev);
                svv = LIMV(((c23v * LVFU(rgb[1][(indx + v1) >> 1]) + c23v * LVFU(rgb[1][(indx + v3) >> 1]) + LVFU(rgb[1][(indx + v5) >> 1]) + LVFU(rgb[1][(indx - v1) >> 1]) + tempv - c32v * LVFU(rgb[0][(indx1 + v1)]) - c8v * LVFU(rgb[0][(indx1 + v2)]))) / c3145680v, zerov, onev);
                //Horizontal and vertical color differences
                tempv = LVFU( rgb[0][indx1] ) / c65535v;
                _mm_storeu_ps( &vdif[indx1], (sgv * nvv + ngv * svv) / (ngv + sgv) - tempv );
                _mm_storeu_ps( &hdif[indx1], (wgv * evv + egv * wvv) / (egv + wgv) - tempv );
            }

            // borders without SSE
            for (; col < width - 5; col += 2, indx += 2, indx1++) {
                //N,E,W,S Gradients
                ng = (eps + (fabsf(rgb[1][(indx - v1) >> 1] - rgb[1][(indx - v3) >> 1]) + fabsf(rgb[0][indx1] - rgb[0][(indx1 - v1)])) / 65535.f);;
                eg = (eps + (fabsf(rgb[1][(indx + h1) >> 1] - rgb[1][(indx + h3) >> 1]) + fabsf(rgb[0][indx1] - rgb[0][(indx1 + h1)])) / 65535.f);
                wg = (eps + (fabsf(rgb[1][(indx - h1) >> 1] - rgb[1][(indx - h3) >> 1]) + fabsf(rgb[0][indx1] - rgb[0][(indx1 - h1)])) / 65535.f);
                sg = (eps + (fabsf(rgb[1][(indx + v1) >> 1] - rgb[1][(indx + v3) >> 1]) + fabsf(rgb[0][indx1] - rgb[0][(indx1 + v1)])) / 65535.f);
                //N,E,W,S High Order Interpolation (Li & Randhawa)
                //N,E,W,S Hamilton Adams Interpolation
                // (48.f * 65535.f) = 3145680.f
                nv = LIM(((23.0f * rgb[1][(indx - v1) >> 1] + 23.0f * rgb[1][(indx - v3) >> 1] + rgb[1][(indx - v5) >> 1] + rgb[1][(indx + v1) >> 1] + 40.0f * rgb[0][indx1] - 32.0f * rgb[0][(indx1 - v1)] - 8.0f * rgb[0][(indx1 - v2)])) / 3145680.f, 0.0f, 1.0f);
                ev = LIM(((23.0f * rgb[1][(indx + h1) >> 1] + 23.0f * rgb[1][(indx + h3) >> 1] + rgb[1][(indx + h5) >> 1] + rgb[1][(indx - h1) >> 1] + 40.0f * rgb[0][indx1] - 32.0f * rgb[0][(indx1 + h1)] - 8.0f * rgb[0][(indx1 + h2)])) / 3145680.f, 0.0f, 1.0f);
                wv = LIM(((23.0f * rgb[1][(indx - h1) >> 1] + 23.0f * rgb[1][(indx - h3) >> 1] + rgb[1][(indx - h5) >> 1] + rgb[1][(indx + h1) >> 1] + 40.0f * rgb[0][indx1] - 32.0f * rgb[0][(indx1 - h1)] - 8.0f * rgb[0][(indx1 - h2)])) / 3145680.f, 0.0f, 1.0f);
                sv = LIM(((23.0f * rgb[1][(indx + v1) >> 1] + 23.0f * rgb[1][(indx + v3) >> 1] + rgb[1][(indx + v5) >> 1] + rgb[1][(indx - v1) >> 1] + 40.0f * rgb[0][indx1] - 32.0f * rgb[0][(indx1 + v1)] - 8.0f * rgb[0][(indx1 + v2)])) / 3145680.f, 0.0f, 1.0f);
                //Horizontal and vertical color differences
                vdif[indx1] = (sg * nv + ng * sv) / (ng + sg) - (rgb[0][indx1]) / 65535.f;
                hdif[indx1] = (wg * ev + eg * wv) / (eg + wg) - (rgb[0][indx1]) / 65535.f;
            }
        }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.26);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++) {
            int col, d, indx1;

            for (col = 7 + (FC(row, 1) & 1), indx1 = (row * width + col) >> 1, d = FC(row, col) / 2; col < width - 14; col += 8, indx1 += 4) {
                //H&V integrated gaussian vector over variance on color differences
                //Mod Jacques 3/2013
                ngv = LIMV(epssqv + c78v * SQRV(LVFU(vdif[indx1])) + c69v * (SQRV(LVFU(vdif[indx1 - v1])) + SQRV(LVFU(vdif[indx1 + v1]))) + c51v * (SQRV(LVFU(vdif[indx1 - v2])) + SQRV(LVFU(vdif[indx1 + v2]))) + c21v * (SQRV(LVFU(vdif[indx1 - v3])) + SQRV(LVFU(vdif[indx1 + v3]))) - c6v * SQRV(LVFU(vdif[indx1 - v1]) + LVFU(vdif[indx1]) + LVFU(vdif[indx1 + v1]))
                           - c10v * (SQRV(LVFU(vdif[indx1 - v2]) + LVFU(vdif[indx1 - v1]) + LVFU(vdif[indx1])) + SQRV(LVFU(vdif[indx1]) + LVFU(vdif[indx1 + v1]) + LVFU(vdif[indx1 + v2]))) - c7v * (SQRV(LVFU(vdif[indx1 - v3]) + LVFU(vdif[indx1 - v2]) + LVFU(vdif[indx1 - v1])) + SQRV(LVFU(vdif[indx1 + v1]) + LVFU(vdif[indx1 + v2]) + LVFU(vdif[indx1 + v3]))), zerov, onev);
                egv = LIMV(epssqv + c78v * SQRV(LVFU(hdif[indx1])) + c69v * (SQRV(LVFU(hdif[indx1 - h1])) + SQRV(LVFU(hdif[indx1 + h1]))) + c51v * (SQRV(LVFU(hdif[indx1 - h2])) + SQRV(LVFU(hdif[indx1 + h2]))) + c21v * (SQRV(LVFU(hdif[indx1 - h3])) + SQRV(LVFU(hdif[indx1 + h3]))) - c6v * SQRV(LVFU(hdif[indx1 - h1]) + LVFU(hdif[indx1]) + LVFU(hdif[indx1 + h1]))
                           - c10v * (SQRV(LVFU(hdif[indx1 - h2]) + LVFU(hdif[indx1 - h1]) + LVFU(hdif[indx1])) + SQRV(LVFU(hdif[indx1]) + LVFU(hdif[indx1 + h1]) + LVFU(hdif[indx1 + h2]))) - c7v * (SQRV(LVFU(hdif[indx1 - h3]) + LVFU(hdif[indx1 - h2]) + LVFU(hdif[indx1 - h1])) + SQRV(LVFU(hdif[indx1 + h1]) + LVFU(hdif[indx1 + h2]) + LVFU(hdif[indx1 + h3]))), zerov, onev);
                //Limit chrominance using H/V neighbourhood
                nvv = median(d725v * LVFU(vdif[indx1]) + d1375v * LVFU(vdif[indx1 - v1]) + d1375v * LVFU(vdif[indx1 + v1]), LVFU(vdif[indx1 - v1]), LVFU(vdif[indx1 + v1]));
                evv = median(d725v * LVFU(hdif[indx1]) + d1375v * LVFU(hdif[indx1 - h1]) + d1375v * LVFU(hdif[indx1 + h1]), LVFU(hdif[indx1 - h1]), LVFU(hdif[indx1 + h1]));
                //Chrominance estimation
                tempv = (egv * nvv + ngv * evv) / (ngv + egv);
                _mm_storeu_ps(&(chr[d][indx1]), tempv);
                //Green channel population
                temp1v = c65535v * tempv + LVFU(rgb[0][indx1]);
                _mm_storeu_ps( &(rgb[0][indx1]), temp1v );
            }

            for (; col < width - 7; col += 2, indx1++) {
                //H&V integrated gaussian vector over variance on color differences
                //Mod Jacques 3/2013
                ng = LIM(epssq + 78.0f * SQR(vdif[indx1]) + 69.0f * (SQR(vdif[indx1 - v1]) + SQR(vdif[indx1 + v1])) + 51.0f * (SQR(vdif[indx1 - v2]) + SQR(vdif[indx1 + v2])) + 21.0f * (SQR(vdif[indx1 - v3]) + SQR(vdif[indx1 + v3])) - 6.0f * SQR(vdif[indx1 - v1] + vdif[indx1] + vdif[indx1 + v1])
                         - 10.0f * (SQR(vdif[indx1 - v2] + vdif[indx1 - v1] + vdif[indx1]) + SQR(vdif[indx1] + vdif[indx1 + v1] + vdif[indx1 + v2])) - 7.0f * (SQR(vdif[indx1 - v3] + vdif[indx1 - v2] + vdif[indx1 - v1]) + SQR(vdif[indx1 + v1] + vdif[indx1 + v2] + vdif[indx1 + v3])), 0.f, 1.f);
                eg = LIM(epssq + 78.0f * SQR(hdif[indx1]) + 69.0f * (SQR(hdif[indx1 - h1]) + SQR(hdif[indx1 + h1])) + 51.0f * (SQR(hdif[indx1 - h2]) + SQR(hdif[indx1 + h2])) + 21.0f * (SQR(hdif[indx1 - h3]) + SQR(hdif[indx1 + h3])) - 6.0f * SQR(hdif[indx1 - h1] + hdif[indx1] + hdif[indx1 + h1])
                         - 10.0f * (SQR(hdif[indx1 - h2] + hdif[indx1 - h1] + hdif[indx1]) + SQR(hdif[indx1] + hdif[indx1 + h1] + hdif[indx1 + h2])) - 7.0f * (SQR(hdif[indx1 - h3] + hdif[indx1 - h2] + hdif[indx1 - h1]) + SQR(hdif[indx1 + h1] + hdif[indx1 + h2] + hdif[indx1 + h3])), 0.f, 1.f);
                //Limit chrominance using H/V neighbourhood
                nv = median(0.725f * vdif[indx1] + 0.1375f * vdif[indx1 - v1] + 0.1375f * vdif[indx1 + v1], vdif[indx1 - v1], vdif[indx1 + v1]);
                ev = median(0.725f * hdif[indx1] + 0.1375f * hdif[indx1 - h1] + 0.1375f * hdif[indx1 + h1], hdif[indx1 - h1], hdif[indx1 + h1]);
                //Chrominance estimation
                chr[d][indx1] = (eg * nv + ng * ev) / (ng + eg);
                //Green channel population
                rgb[0][indx1] = rgb[0][indx1] + 65535.f * chr[d][indx1];
            }
        }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.39);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++) {
            int col, indx, c;

            for (col = 7 + (FC(row, 1) & 1), indx = row * width + col, c = 1 - FC(row, col) / 2; col < width - 14; col += 8, indx += 8) {
                //NW,NE,SW,SE Gradients
                nwgv = onev / (epsv + vabsf(LVFU(chr[c][(indx - v1 - h1) >> 1]) - LVFU(chr[c][(indx - v3 - h3) >> 1])) + vabsf(LVFU(chr[c][(indx + v1 + h1) >> 1]) - LVFU(chr[c][(indx - v3 - h3) >> 1])));
                negv = onev / (epsv + vabsf(LVFU(chr[c][(indx - v1 + h1) >> 1]) - LVFU(chr[c][(indx - v3 + h3) >> 1])) + vabsf(LVFU(chr[c][(indx + v1 - h1) >> 1]) - LVFU(chr[c][(indx - v3 + h3) >> 1])));
                swgv = onev / (epsv + vabsf(LVFU(chr[c][(indx + v1 - h1) >> 1]) - LVFU(chr[c][(indx + v3 + h3) >> 1])) + vabsf(LVFU(chr[c][(indx - v1 + h1) >> 1]) - LVFU(chr[c][(indx + v3 - h3) >> 1])));
                segv = onev / (epsv + vabsf(LVFU(chr[c][(indx + v1 + h1) >> 1]) - LVFU(chr[c][(indx + v3 - h3) >> 1])) + vabsf(LVFU(chr[c][(indx - v1 - h1) >> 1]) - LVFU(chr[c][(indx + v3 + h3) >> 1])));
                //Limit NW,NE,SW,SE Color differences
                nwvv = median(LVFU(chr[c][(indx - v1 - h1) >> 1]), LVFU(chr[c][(indx - v3 - h1) >> 1]), LVFU(chr[c][(indx - v1 - h3) >> 1]));
                nevv = median(LVFU(chr[c][(indx - v1 + h1) >> 1]), LVFU(chr[c][(indx - v3 + h1) >> 1]), LVFU(chr[c][(indx - v1 + h3) >> 1]));
                swvv = median(LVFU(chr[c][(indx + v1 - h1) >> 1]), LVFU(chr[c][(indx + v3 - h1) >> 1]), LVFU(chr[c][(indx + v1 - h3) >> 1]));
                sevv = median(LVFU(chr[c][(indx + v1 + h1) >> 1]), LVFU(chr[c][(indx + v3 + h1) >> 1]), LVFU(chr[c][(indx + v1 + h3) >> 1]));
                //Interpolate chrominance: R@B and B@R
                tempv = (nwgv * nwvv + negv * nevv + swgv * swvv + segv * sevv) / (nwgv + negv + swgv + segv);
                _mm_storeu_ps( &(chr[c][indx >> 1]), tempv);
            }

            for (; col < width - 7; col += 2, indx += 2) {
                //NW,NE,SW,SE Gradients
                nwg = 1.0f / (eps + fabsf(chr[c][(indx - v1 - h1) >> 1] - chr[c][(indx - v3 - h3) >> 1]) + fabsf(chr[c][(indx + v1 + h1) >> 1] - chr[c][(indx - v3 - h3) >> 1]));
                neg = 1.0f / (eps + fabsf(chr[c][(indx - v1 + h1) >> 1] - chr[c][(indx - v3 + h3) >> 1]) + fabsf(chr[c][(indx + v1 - h1) >> 1] - chr[c][(indx - v3 + h3) >> 1]));
                swg = 1.0f / (eps + fabsf(chr[c][(indx + v1 - h1) >> 1] - chr[c][(indx + v3 + h3) >> 1]) + fabsf(chr[c][(indx - v1 + h1) >> 1] - chr[c][(indx + v3 - h3) >> 1]));
                seg = 1.0f / (eps + fabsf(chr[c][(indx + v1 + h1) >> 1] - chr[c][(indx + v3 - h3) >> 1]) + fabsf(chr[c][(indx - v1 - h1) >> 1] - chr[c][(indx + v3 + h3) >> 1]));
                //Limit NW,NE,SW,SE Color differences
                nwv = median(chr[c][(indx - v1 - h1) >> 1], chr[c][(indx - v3 - h1) >> 1], chr[c][(indx - v1 - h3) >> 1]);
                nev = median(chr[c][(indx - v1 + h1) >> 1], chr[c][(indx - v3 + h1) >> 1], chr[c][(indx - v1 + h3) >> 1]);
                swv = median(chr[c][(indx + v1 - h1) >> 1], chr[c][(indx + v3 - h1) >> 1], chr[c][(indx + v1 - h3) >> 1]);
                sev = median(chr[c][(indx + v1 + h1) >> 1], chr[c][(indx + v3 + h1) >> 1], chr[c][(indx + v1 + h3) >> 1]);
                //Interpolate chrominance: R@B and B@R
                chr[c][indx >> 1] = (nwg * nwv + neg * nev + swg * swv + seg * sev) / (nwg + neg + swg + seg);
            }
        }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.65);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++) {
            int col, indx;

            for (col = 7 + (FC(row, 0) & 1), indx = row * width + col; col < width - 14; col += 8, indx += 8) {
                //N,E,W,S Gradients
                ngv = onev / (epsv + vabsf(LVFU(chr[0][(indx - v1) >> 1]) - LVFU(chr[0][(indx - v3) >> 1])) + vabsf(LVFU(chr[0][(indx + v1) >> 1]) - LVFU(chr[0][(indx - v3) >> 1])));
                egv = onev / (epsv + vabsf(LVFU(chr[0][(indx + h1) >> 1]) - LVFU(chr[0][(indx + h3) >> 1])) + vabsf(LVFU(chr[0][(indx - h1) >> 1]) - LVFU(chr[0][(indx + h3) >> 1])));
                wgv = onev / (epsv + vabsf(LVFU(chr[0][(indx - h1) >> 1]) - LVFU(chr[0][(indx - h3) >> 1])) + vabsf(LVFU(chr[0][(indx + h1) >> 1]) - LVFU(chr[0][(indx - h3) >> 1])));
                sgv = onev / (epsv + vabsf(LVFU(chr[0][(indx + v1) >> 1]) - LVFU(chr[0][(indx + v3) >> 1])) + vabsf(LVFU(chr[0][(indx - v1) >> 1]) - LVFU(chr[0][(indx + v3) >> 1])));
                //Interpolate chrominance: R@G and B@G
                tempv = ((ngv * LVFU(chr[0][(indx - v1) >> 1]) + egv * LVFU(chr[0][(indx + h1) >> 1]) + wgv * LVFU(chr[0][(indx - h1) >> 1]) + sgv * LVFU(chr[0][(indx + v1) >> 1])) / (ngv + egv + wgv + sgv));
                _mm_storeu_ps( &chr[0 + 2][indx >> 1], tempv);
            }

            for (; col < width - 7; col += 2, indx += 2) {
                //N,E,W,S Gradients
                ng = 1.0f / (eps + fabsf(chr[0][(indx - v1) >> 1] - chr[0][(indx - v3) >> 1]) + fabsf(chr[0][(indx + v1) >> 1] - chr[0][(indx - v3) >> 1]));
                eg = 1.0f / (eps + fabsf(chr[0][(indx + h1) >> 1] - chr[0][(indx + h3) >> 1]) + fabsf(chr[0][(indx - h1) >> 1] - chr[0][(indx + h3) >> 1]));
                wg = 1.0f / (eps + fabsf(chr[0][(indx - h1) >> 1] - chr[0][(indx - h3) >> 1]) + fabsf(chr[0][(indx + h1) >> 1] - chr[0][(indx - h3) >> 1]));
                sg = 1.0f / (eps + fabsf(chr[0][(indx + v1) >> 1] - chr[0][(indx + v3) >> 1]) + fabsf(chr[0][(indx - v1) >> 1] - chr[0][(indx + v3) >> 1]));
                //Interpolate chrominance: R@G and B@G
                chr[0 + 2][indx >> 1] = ((ng * chr[0][(indx - v1) >> 1] + eg * chr[0][(indx + h1) >> 1] + wg * chr[0][(indx - h1) >> 1] + sg * chr[0][(indx + v1) >> 1]) / (ng + eg + wg + sg));
            }
        }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.78);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++) {
            int col, indx;

            for (col = 7 + (FC(row, 0) & 1), indx = row * width + col; col < width - 14; col += 8, indx += 8) {
                //N,E,W,S Gradients
                ngv = onev / (epsv + vabsf(LVFU(chr[1][(indx - v1) >> 1]) - LVFU(chr[1][(indx - v3) >> 1])) + vabsf(LVFU(chr[1][(indx + v1) >> 1]) - LVFU(chr[1][(indx - v3) >> 1])));
                egv = onev / (epsv + vabsf(LVFU(chr[1][(indx + h1) >> 1]) - LVFU(chr[1][(indx + h3) >> 1])) + vabsf(LVFU(chr[1][(indx - h1) >> 1]) - LVFU(chr[1][(indx + h3) >> 1])));
                wgv = onev / (epsv + vabsf(LVFU(chr[1][(indx - h1) >> 1]) - LVFU(chr[1][(indx - h3) >> 1])) + vabsf(LVFU(chr[1][(indx + h1) >> 1]) - LVFU(chr[1][(indx - h3) >> 1])));
                sgv = onev / (epsv + vabsf(LVFU(chr[1][(indx + v1) >> 1]) - LVFU(chr[1][(indx + v3) >> 1])) + vabsf(LVFU(chr[1][(indx - v1) >> 1]) - LVFU(chr[1][(indx + v3) >> 1])));
                //Interpolate chrominance: R@G and B@G
                tempv = ((ngv * LVFU(chr[1][(indx - v1) >> 1]) + egv * LVFU(chr[1][(indx + h1) >> 1]) + wgv * LVFU(chr[1][(indx - h1) >> 1]) + sgv * LVFU(chr[1][(indx + v1) >> 1])) / (ngv + egv + wgv + sgv));
                _mm_storeu_ps( &chr[1 + 2][indx >> 1], tempv);
            }

            for (; col < width - 7; col += 2, indx += 2) {
                //N,E,W,S Gradients
                ng = 1.0f / (eps + fabsf(chr[1][(indx - v1) >> 1] - chr[1][(indx - v3) >> 1]) + fabsf(chr[1][(indx + v1) >> 1] - chr[1][(indx - v3) >> 1]));
                eg = 1.0f / (eps + fabsf(chr[1][(indx + h1) >> 1] - chr[1][(indx + h3) >> 1]) + fabsf(chr[1][(indx - h1) >> 1] - chr[1][(indx + h3) >> 1]));
                wg = 1.0f / (eps + fabsf(chr[1][(indx - h1) >> 1] - chr[1][(indx - h3) >> 1]) + fabsf(chr[1][(indx + h1) >> 1] - chr[1][(indx - h3) >> 1]));
                sg = 1.0f / (eps + fabsf(chr[1][(indx + v1) >> 1] - chr[1][(indx + v3) >> 1]) + fabsf(chr[1][(indx - v1) >> 1] - chr[1][(indx + v3) >> 1]));
                //Interpolate chrominance: R@G and B@G
                chr[1 + 2][indx >> 1] = ((ng * chr[1][(indx - v1) >> 1] + eg * chr[1][(indx + h1) >> 1] + wg * chr[1][(indx - h1) >> 1] + sg * chr[1][(indx + v1) >> 1]) / (ng + eg + wg + sg));
            }
        }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.91);
            }
        }
        float *src1, *src2, *redsrc0, *redsrc1, *bluesrc0, *bluesrc1;
#ifdef _OPENMP
        #pragma omp for
#endif

        for(int row = 7; row < height - 7; row++) {
            int col, indx, fc;
            fc = FC(row, 7) & 1;
            src1 = rgb[fc];
            src2 = rgb[fc ^ 1];
            redsrc0 = chr[fc << 1];
            redsrc1 = chr[(fc ^ 1) << 1];
            bluesrc0 = chr[(fc << 1) + 1];
            bluesrc1 = chr[((fc ^ 1) << 1) + 1];

            for(col = 7, indx = row * width + col; col < width - 14; col += 8, indx += 8) {
                temp1v = LVFU( src1[indx >> 1] );
                temp2v = LVFU( src2[(indx + 1) >> 1] );
                tempv = _mm_shuffle_ps( temp1v, temp2v, _MM_SHUFFLE( 1, 0, 1, 0 ) );
                tempv = _mm_shuffle_ps( tempv, tempv, _MM_SHUFFLE( 3, 1, 2, 0 ) );
                _mm_storeu_ps( &green[row][col], CLIPV( tempv ));
                temp5v = LVFU(redsrc0[indx >> 1]);
                temp6v = LVFU(redsrc1[(indx + 1) >> 1]);
                temp3v = _mm_shuffle_ps( temp5v, temp6v, _MM_SHUFFLE( 1, 0, 1, 0 ) );
                temp3v = _mm_shuffle_ps( temp3v, temp3v, _MM_SHUFFLE( 3, 1, 2, 0 ) );
                temp3v = CLIPV( tempv - c65535v * temp3v );
                _mm_storeu_ps( &red[row][col], temp3v);
                temp7v = LVFU(bluesrc0[indx >> 1]);
                temp8v = LVFU(bluesrc1[(indx + 1) >> 1]);
                temp4v = _mm_shuffle_ps( temp7v, temp8v, _MM_SHUFFLE( 1, 0, 1, 0 ) );
                temp4v = _mm_shuffle_ps( temp4v, temp4v, _MM_SHUFFLE( 3, 1, 2, 0 ) );
                temp4v = CLIPV( tempv - c65535v * temp4v );
                _mm_storeu_ps( &blue[row][col], temp4v);

                tempv = _mm_shuffle_ps( temp1v, temp2v, _MM_SHUFFLE( 3, 2, 3, 2 ) );
                tempv = _mm_shuffle_ps( tempv, tempv, _MM_SHUFFLE( 3, 1, 2, 0 ) );
                _mm_storeu_ps( &green[row][col + 4], CLIPV( tempv ));

                temp3v = _mm_shuffle_ps( temp5v, temp6v, _MM_SHUFFLE( 3, 2, 3, 2 ) );
                temp3v = _mm_shuffle_ps( temp3v, temp3v, _MM_SHUFFLE( 3, 1, 2, 0 ) );
                temp3v = CLIPV( tempv - c65535v * temp3v );
                _mm_storeu_ps( &red[row][col + 4], temp3v);
                temp4v = _mm_shuffle_ps( temp7v, temp8v, _MM_SHUFFLE( 3, 2, 3, 2 ) );
                temp4v = _mm_shuffle_ps( temp4v, temp4v, _MM_SHUFFLE( 3, 1, 2, 0 ) );
                temp4v = CLIPV( tempv - c65535v * temp4v );
                _mm_storeu_ps( &blue[row][col + 4], temp4v);
            }

            for(; col < width - 7; col++, indx += 2) {
                red  [row][col] = CLIP(src1[indx >> 1] - 65535.f * redsrc0[indx >> 1]);
                green[row][col] = CLIP(src1[indx >> 1]);
                blue [row][col] = CLIP(src1[indx >> 1] - 65535.f * bluesrc0[indx >> 1]);
                col++;
                red  [row][col] = CLIP(src2[(indx + 1) >> 1] - 65535.f * redsrc1[(indx + 1) >> 1]);
                green[row][col] = CLIP(src2[(indx + 1) >> 1]);
                blue [row][col] = CLIP(src2[(indx + 1) >> 1] - 65535.f * bluesrc1[(indx + 1) >> 1]);
            }
        }
    }// End of parallelization

    if (plistener) {
        plistener->setProgress (1.0);
    }

    free(chrarray);
    free(rgbarray);
    free(vdif);
    free(hdif);
}
#undef CLIPV
#else
void RawImageSource::igv_interpolate(int winw, int winh)
{
    static const float eps = 1e-5f, epssq = 1e-5f; //mod epssq -10f =>-5f Jacques 3/2013 to prevent artifact (divide by zero)
    static const int h1 = 1, h2 = 2, h3 = 3, h4 = 4, h5 = 5, h6 = 6;
    const int width = winw, height = winh;
    const int v1 = 1 * width, v2 = 2 * width, v3 = 3 * width, v4 = 4 * width, v5 = 5 * width, v6 = 6 * width;
    float* rgb[3];
    float* chr[2];
    float (*rgbarray), *vdif, *hdif, (*chrarray);

    rgbarray    = (float (*)) calloc(width * height * 3, sizeof( float));
    rgb[0] = rgbarray;
    rgb[1] = rgbarray + (width * height);
    rgb[2] = rgbarray + 2 * (width * height);

    chrarray    = (float (*)) calloc(width * height * 2, sizeof( float));
    chr[0] = chrarray;
    chr[1] = chrarray + (width * height);

    vdif  = (float (*))    calloc(width * height / 2, sizeof * vdif);
    hdif  = (float (*))    calloc(width * height / 2, sizeof * hdif);

    border_interpolate2(winw, winh, 7, rawData, red, green, blue);

    if (plistener) {
        plistener->setProgressStr (Glib::ustring::compose(M("TP_RAW_DMETHOD_PROGRESSBAR"), RAWParams::BayerSensor::getMethodString(RAWParams::BayerSensor::Method::IGV)));
        plistener->setProgress (0.0);
    }

#ifdef _OPENMP
    #pragma omp parallel default(none) shared(rgb,vdif,hdif,chr)
#endif
    {

        float ng, eg, wg, sg, nv, ev, wv, sv, nwg, neg, swg, seg, nwv, nev, swv, sev;

#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 0; row < height - 0; row++)
            for (int col = 0, indx = row * width + col; col < width - 0; col++, indx++) {
                int c = FC(row, col);
                rgb[c][indx] = CLIP(rawData[row][col]); //rawData = RT datas
            }

//  border_interpolate2(7, rgb);

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.13);
            }
        }

#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 5; row < height - 5; row++)
            for (int col = 5 + (FC(row, 1) & 1), indx = row * width + col, c = FC(row, col); col < width - 5; col += 2, indx += 2) {
                //N,E,W,S Gradients
                ng = (eps + (fabsf(rgb[1][indx - v1] - rgb[1][indx - v3]) + fabsf(rgb[c][indx] - rgb[c][indx - v2])) / 65535.f);;
                eg = (eps + (fabsf(rgb[1][indx + h1] - rgb[1][indx + h3]) + fabsf(rgb[c][indx] - rgb[c][indx + h2])) / 65535.f);
                wg = (eps + (fabsf(rgb[1][indx - h1] - rgb[1][indx - h3]) + fabsf(rgb[c][indx] - rgb[c][indx - h2])) / 65535.f);
                sg = (eps + (fabsf(rgb[1][indx + v1] - rgb[1][indx + v3]) + fabsf(rgb[c][indx] - rgb[c][indx + v2])) / 65535.f);
                //N,E,W,S High Order Interpolation (Li & Randhawa)
                //N,E,W,S Hamilton Adams Interpolation
                // (48.f * 65535.f) = 3145680.f
                nv = LIM(((23.0f * rgb[1][indx - v1] + 23.0f * rgb[1][indx - v3] + rgb[1][indx - v5] + rgb[1][indx + v1] + 40.0f * rgb[c][indx] - 32.0f * rgb[c][indx - v2] - 8.0f * rgb[c][indx - v4])) / 3145680.f, 0.0f, 1.0f);
                ev = LIM(((23.0f * rgb[1][indx + h1] + 23.0f * rgb[1][indx + h3] + rgb[1][indx + h5] + rgb[1][indx - h1] + 40.0f * rgb[c][indx] - 32.0f * rgb[c][indx + h2] - 8.0f * rgb[c][indx + h4])) / 3145680.f, 0.0f, 1.0f);
                wv = LIM(((23.0f * rgb[1][indx - h1] + 23.0f * rgb[1][indx - h3] + rgb[1][indx - h5] + rgb[1][indx + h1] + 40.0f * rgb[c][indx] - 32.0f * rgb[c][indx - h2] - 8.0f * rgb[c][indx - h4])) / 3145680.f, 0.0f, 1.0f);
                sv = LIM(((23.0f * rgb[1][indx + v1] + 23.0f * rgb[1][indx + v3] + rgb[1][indx + v5] + rgb[1][indx - v1] + 40.0f * rgb[c][indx] - 32.0f * rgb[c][indx + v2] - 8.0f * rgb[c][indx + v4])) / 3145680.f, 0.0f, 1.0f);
                //Horizontal and vertical color differences
                vdif[indx >> 1] = (sg * nv + ng * sv) / (ng + sg) - (rgb[c][indx]) / 65535.f;
                hdif[indx >> 1] = (wg * ev + eg * wv) / (eg + wg) - (rgb[c][indx]) / 65535.f;
            }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.26);
            }
        }

#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++)
            for (int col = 7 + (FC(row, 1) & 1), indx = row * width + col, c = FC(row, col), d = c / 2; col < width - 7; col += 2, indx += 2) {
                //H&V integrated gaussian vector over variance on color differences
                //Mod Jacques 3/2013
                ng = LIM(epssq + 78.0f * SQR(vdif[indx >> 1]) + 69.0f * (SQR(vdif[(indx - v2) >> 1]) + SQR(vdif[(indx + v2) >> 1])) + 51.0f * (SQR(vdif[(indx - v4) >> 1]) + SQR(vdif[(indx + v4) >> 1])) + 21.0f * (SQR(vdif[(indx - v6) >> 1]) + SQR(vdif[(indx + v6) >> 1])) - 6.0f * SQR(vdif[(indx - v2) >> 1] + vdif[indx >> 1] + vdif[(indx + v2) >> 1])
                         - 10.0f * (SQR(vdif[(indx - v4) >> 1] + vdif[(indx - v2) >> 1] + vdif[indx >> 1]) + SQR(vdif[indx >> 1] + vdif[(indx + v2) >> 1] + vdif[(indx + v4) >> 1])) - 7.0f * (SQR(vdif[(indx - v6) >> 1] + vdif[(indx - v4) >> 1] + vdif[(indx - v2) >> 1]) + SQR(vdif[(indx + v2) >> 1] + vdif[(indx + v4) >> 1] + vdif[(indx + v6) >> 1])), 0.f, 1.f);
                eg = LIM(epssq + 78.0f * SQR(hdif[indx >> 1]) + 69.0f * (SQR(hdif[(indx - h2) >> 1]) + SQR(hdif[(indx + h2) >> 1])) + 51.0f * (SQR(hdif[(indx - h4) >> 1]) + SQR(hdif[(indx + h4) >> 1])) + 21.0f * (SQR(hdif[(indx - h6) >> 1]) + SQR(hdif[(indx + h6) >> 1])) - 6.0f * SQR(hdif[(indx - h2) >> 1] + hdif[indx >> 1] + hdif[(indx + h2) >> 1])
                         - 10.0f * (SQR(hdif[(indx - h4) >> 1] + hdif[(indx - h2) >> 1] + hdif[indx >> 1]) + SQR(hdif[indx >> 1] + hdif[(indx + h2) >> 1] + hdif[(indx + h4) >> 1])) - 7.0f * (SQR(hdif[(indx - h6) >> 1] + hdif[(indx - h4) >> 1] + hdif[(indx - h2) >> 1]) + SQR(hdif[(indx + h2) >> 1] + hdif[(indx + h4) >> 1] + hdif[(indx + h6) >> 1])), 0.f, 1.f);
                //Limit chrominance using H/V neighbourhood
                nv = median(0.725f * vdif[indx >> 1] + 0.1375f * vdif[(indx - v2) >> 1] + 0.1375f * vdif[(indx + v2) >> 1], vdif[(indx - v2) >> 1], vdif[(indx + v2) >> 1]);
                ev = median(0.725f * hdif[indx >> 1] + 0.1375f * hdif[(indx - h2) >> 1] + 0.1375f * hdif[(indx + h2) >> 1], hdif[(indx - h2) >> 1], hdif[(indx + h2) >> 1]);
                //Chrominance estimation
                chr[d][indx] = (eg * nv + ng * ev) / (ng + eg);
                //Green channel population
                rgb[1][indx] = rgb[c][indx] + 65535.f * chr[d][indx];
            }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.39);
            }
        }

//  free(vdif); free(hdif);
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row += 2)
            for (int col = 7 + (FC(row, 1) & 1), indx = row * width + col, c = 1 - FC(row, col) / 2; col < width - 7; col += 2, indx += 2) {
                //NW,NE,SW,SE Gradients
                nwg = 1.0f / (eps + fabsf(chr[c][indx - v1 - h1] - chr[c][indx - v3 - h3]) + fabsf(chr[c][indx + v1 + h1] - chr[c][indx - v3 - h3]));
                neg = 1.0f / (eps + fabsf(chr[c][indx - v1 + h1] - chr[c][indx - v3 + h3]) + fabsf(chr[c][indx + v1 - h1] - chr[c][indx - v3 + h3]));
                swg = 1.0f / (eps + fabsf(chr[c][indx + v1 - h1] - chr[c][indx + v3 + h3]) + fabsf(chr[c][indx - v1 + h1] - chr[c][indx + v3 - h3]));
                seg = 1.0f / (eps + fabsf(chr[c][indx + v1 + h1] - chr[c][indx + v3 - h3]) + fabsf(chr[c][indx - v1 - h1] - chr[c][indx + v3 + h3]));
                //Limit NW,NE,SW,SE Color differences
                nwv = median(chr[c][indx - v1 - h1], chr[c][indx - v3 - h1], chr[c][indx - v1 - h3]);
                nev = median(chr[c][indx - v1 + h1], chr[c][indx - v3 + h1], chr[c][indx - v1 + h3]);
                swv = median(chr[c][indx + v1 - h1], chr[c][indx + v3 - h1], chr[c][indx + v1 - h3]);
                sev = median(chr[c][indx + v1 + h1], chr[c][indx + v3 + h1], chr[c][indx + v1 + h3]);
                //Interpolate chrominance: R@B and B@R
                chr[c][indx] = (nwg * nwv + neg * nev + swg * swv + seg * sev) / (nwg + neg + swg + seg);
            }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.52);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 8; row < height - 7; row += 2)
            for (int col = 7 + (FC(row, 1) & 1), indx = row * width + col, c = 1 - FC(row, col) / 2; col < width - 7; col += 2, indx += 2) {
                //NW,NE,SW,SE Gradients
                nwg = 1.0f / (eps + fabsf(chr[c][indx - v1 - h1] - chr[c][indx - v3 - h3]) + fabsf(chr[c][indx + v1 + h1] - chr[c][indx - v3 - h3]));
                neg = 1.0f / (eps + fabsf(chr[c][indx - v1 + h1] - chr[c][indx - v3 + h3]) + fabsf(chr[c][indx + v1 - h1] - chr[c][indx - v3 + h3]));
                swg = 1.0f / (eps + fabsf(chr[c][indx + v1 - h1] - chr[c][indx + v3 + h3]) + fabsf(chr[c][indx - v1 + h1] - chr[c][indx + v3 - h3]));
                seg = 1.0f / (eps + fabsf(chr[c][indx + v1 + h1] - chr[c][indx + v3 - h3]) + fabsf(chr[c][indx - v1 - h1] - chr[c][indx + v3 + h3]));
                //Limit NW,NE,SW,SE Color differences
                nwv = median(chr[c][indx - v1 - h1], chr[c][indx - v3 - h1], chr[c][indx - v1 - h3]);
                nev = median(chr[c][indx - v1 + h1], chr[c][indx - v3 + h1], chr[c][indx - v1 + h3]);
                swv = median(chr[c][indx + v1 - h1], chr[c][indx + v3 - h1], chr[c][indx + v1 - h3]);
                sev = median(chr[c][indx + v1 + h1], chr[c][indx + v3 + h1], chr[c][indx + v1 + h3]);
                //Interpolate chrominance: R@B and B@R
                chr[c][indx] = (nwg * nwv + neg * nev + swg * swv + seg * sev) / (nwg + neg + swg + seg);
            }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.65);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++)
            for (int col = 7 + (FC(row, 0) & 1), indx = row * width + col; col < width - 7; col += 2, indx += 2) {
                //N,E,W,S Gradients
                ng = 1.0f / (eps + fabsf(chr[0][indx - v1] - chr[0][indx - v3]) + fabsf(chr[0][indx + v1] - chr[0][indx - v3]));
                eg = 1.0f / (eps + fabsf(chr[0][indx + h1] - chr[0][indx + h3]) + fabsf(chr[0][indx - h1] - chr[0][indx + h3]));
                wg = 1.0f / (eps + fabsf(chr[0][indx - h1] - chr[0][indx - h3]) + fabsf(chr[0][indx + h1] - chr[0][indx - h3]));
                sg = 1.0f / (eps + fabsf(chr[0][indx + v1] - chr[0][indx + v3]) + fabsf(chr[0][indx - v1] - chr[0][indx + v3]));
                //Interpolate chrominance: R@G and B@G
                chr[0][indx] = ((ng * chr[0][indx - v1] + eg * chr[0][indx + h1] + wg * chr[0][indx - h1] + sg * chr[0][indx + v1]) / (ng + eg + wg + sg));
            }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.78);
            }
        }
#ifdef _OPENMP
        #pragma omp for
#endif

        for (int row = 7; row < height - 7; row++)
            for (int col = 7 + (FC(row, 0) & 1), indx = row * width + col; col < width - 7; col += 2, indx += 2) {

                //N,E,W,S Gradients
                ng = 1.0f / (eps + fabsf(chr[1][indx - v1] - chr[1][indx - v3]) + fabsf(chr[1][indx + v1] - chr[1][indx - v3]));
                eg = 1.0f / (eps + fabsf(chr[1][indx + h1] - chr[1][indx + h3]) + fabsf(chr[1][indx - h1] - chr[1][indx + h3]));
                wg = 1.0f / (eps + fabsf(chr[1][indx - h1] - chr[1][indx - h3]) + fabsf(chr[1][indx + h1] - chr[1][indx - h3]));
                sg = 1.0f / (eps + fabsf(chr[1][indx + v1] - chr[1][indx + v3]) + fabsf(chr[1][indx - v1] - chr[1][indx + v3]));
                //Interpolate chrominance: R@G and B@G
                chr[1][indx] = ((ng * chr[1][indx - v1] + eg * chr[1][indx + h1] + wg * chr[1][indx - h1] + sg * chr[1][indx + v1]) / (ng + eg + wg + sg));
            }

#ifdef _OPENMP
        #pragma omp single
#endif
        {
            if (plistener) {
                plistener->setProgress (0.91);
            }

            //Interpolate borders
//  border_interpolate2(7, rgb);
        }
        /*
#ifdef _OPENMP
        #pragma omp for
#endif
            for (int row=0; row < height; row++)  //borders
                for (int col=0; col < width; col++) {
                    if (col==7 && row >= 7 && row < height-7)
                        col = width-7;
                    int indxc=row*width+col;
                    red  [row][col] = rgb[indxc][0];
                    green[row][col] = rgb[indxc][1];
                    blue [row][col] = rgb[indxc][2];
                }
        */

#ifdef _OPENMP
        #pragma omp for
#endif

        for(int row = 7; row < height - 7; row++)
            for(int col = 7, indx = row * width + col; col < width - 7; col++, indx++) {
                red  [row][col] = CLIP(rgb[1][indx] - 65535.f * chr[0][indx]);
                green[row][col] = CLIP(rgb[1][indx]);
                blue [row][col] = CLIP(rgb[1][indx] - 65535.f * chr[1][indx]);
            }
    }// End of parallelization

    if (plistener) {
        plistener->setProgress (1.0);
    }

    free(chrarray);
    free(rgbarray);
    free(vdif);
    free(hdif);
}
#endif


@50mm1200s Thanks! I had to denoise some parts. I wish I could use anamorphic! I'll definitely get adapters once I can afford them.

Yeah, they add so much personality to the work, it's incredible. But, be careful on what you get. I had a Kowa 16D once, it was good, but very difficult to use and sometimes you get nasty optical issues, like vignette. I think the best you can buy today as entry-level is the Anamorphot-50, but I would suggest you do your own oval aperture filter, just to see if you adapt to the stretched bokeh aesthetics. I did mines for less than $5 using wood:

Well done. Everything is fine, technically speaking. Not much noise, sharpen is nice, color grading, etc.
I think, though, these kind of videos would benefit from anamorphics and strange transitions, like match cuts or the "passing by" effect.
All and all, very good.

I'd like that too. I can only hide it through Global Draw Off.

Another one, but using ANN (artificial neural network):
https://www.dcl.hpi.uni-potsdam.de/papers/papers/heinze-joint-multiframe.pdf

Demonstration compared to ACR (Adobe Camera Raw). Check the shadows and midtone noise on the car:

[I wonder why no one bother to unite all these research for practical use.]

Another one. The idea of doing pre-demosaicing filtering seems to have spread these last years. This other paper presents denoising before demosaicing using machine-learning (Joint Demosaicing and Denoising via Learned Non-parametric Random Fields):
http://www.nowozin.net/sebastian/papers/khashabi2014demosaicing.pdf

The noise from CMOS or CCD generates interpolation errors (specially "read noise" and "shot noise", page 3), from what I've read on the paper. So, prior to demosaicing, they apply denoising. Below is a demonstration from the paper.  The RTF (Regression Tree Field) is their work, compared to CS (Contour Stencils), NLM (Non-Local Means), MBI (Matlab demosaic function) and the ground truth. Click to enlarge:




And this other paper presents a sharpening/debluring (using deconvolution) before demosaicing. This work is based on the same approach on the superresolution paper shared above (using MAP framework, if I understood right):
https://hal.archives-ouvertes.fr/hal-00408220/document

Demonstration (click to enlarge):




I wonder why no one bother to unite all these research for practical use.

Those gimbal stuff is pretty neat. I'm about to buy a Glidecam, but this Zhiyun is so good and easy to use that I'm considering it. My only concern (besides money) is the weight. I uses old manual lenses and they are pretty heavy. The set is probably 3kg...
Thanks for sharing you experience with it.

Since you guys are talking about demosaicing again, thought I'd share this paper (Multi-Frame Demosaicing and Super-Resolution of Color Images):
http://people.duke.edu/~sf59/TIP_Demos_Final_Color.pdf

To be fair, I don't understand the math, but the method proposed uses super-resolution as a demosaicing algorithm and reconstitures not just luma, but also chroma. From what I understood, he uses low-res versions of the image as "training" of the SR method. Here's the abstract:

In the last two decades, two related categories of problems have been studied independently in the image restoration literature: super-resolution and demosaicing. A closer look at these problems reveals the relation between them, and as conventional color digital cameras suffer from both low-spatial resolution and color-filtering, it is reasonable to address them in a unified context. In this paper, we propose a fast and robust hybrid method of super-resolution and demosaicing, based on a MAP estimation technique by minimizing a multi-term cost function. The L1 norm is used for measuring the difference between the projected estimate of the high-resolution image and each low-resolution image, removing outliers in the data and errors due to possibly inaccurate motion estimation. Bilateral regularization is used for spatially regularizing the luminance component, resulting in sharp edges and forcing interpolation along the edges and not across them. Simultaneously, Tikhonov regularization is used to smooth the
chrominance components. Finally, an additional regularization term is used to force similar edge location and orientation in different color channels. We show that the minimization of the total cost function is relatively easy and fast. Experimental results on synthetic and real data sets confirm the effectiveness of our method.

And here's the practical example (click to enlarge):




The other paper shared some time ago combined sharpen inside demosaicing, this one combines SR. Maybe both ideas could be used together... humm.

At first I thought this was bullshit, because in 7 years working with photography I have never even heard of it, but then found this article from Zeiss:
https://lenspire.zeiss.com/photo/en/article/micro-contrast-and-the-zeiss-pop-by-lloyd-chambers/

Thought it would be nice to share, since we have been discussing lenses on this other thread lately.

BTW, @IDA_ML is right about the ND filter. Set to 24fps, 1/48 shutter speed, open the lens and then fix the exposure with a ND. This will give better motion blur and a better "feel"...
A polarizer can help too, to recover some highlights. I heard the B+W circular polarizers are pretty good.

I'm a bit disappointed of youtube's compression and I'd really love to get some tips from u guys on how to get the best uploads settings for YT. This time I uploaded the 6gB file in DNxHR but I'm still getting a poor result.

I've seen people upscaling it to 4K but isn't there any other workaround ? Please help me guys because my DNxHR file looks way better than the uploaded one and I would love to share it with you.

Don't upscale, unless it's necessary to fit the specs on the platform you're using. On YouTube, the accepted sizes are always 16:9, on HD, FullHD or UHD standards.
On your case, you uploaded using a aspect ratio of 2.35:1 (1920x818px). You should have made a letterbox to reach the 1920x1080px before uploading.
As about the format, see this here:
https://support.google.com/youtube/answer/7126552?hl=en

I would go for this, but haven't tested yet: 1920x1080px, 24fps, ProRes 444, 10-bit, Rec.2020 and audio using FLAC format (will have to use MKV container).

Moving trees is perfect for the ffmpeg filter:
without
with (3 pass + unsharp mask)

Very cool. I think, though, the unsharp filter have has too big ratio. What about this (copied from @Danne post - ratio 3 instead of 5):

  moireeFilter.append( QString( "unsharp=3:3:0.6:3:3:0," ) );

Or maybe no unsharp at all?

@50mm1200s - My name is Fort as in strong in French not Ford like the car. Been dealing with that all my life. 

Oh, sorry about that.

The two biggest issues when shooting video with DSLRs are aliasing artifacts and rolling shutter. Here's a challenge for this app, something to reduce the "Jello" effect. That should be doable, though trying to fix propellers, that's a bit more challenging.

I think the VAF does a decent job on aliasing. About the shutter, this solution is neat:

Thought about this paper clip icon (Resolution of 46x42px, 72dpi, png optimized with pngquant and ect):