ffmpeg sws_scale got distorted image from YUV420P to RGB24

Question

I got distorted image when try to convert YUV420p to RGB24 using sws_scale.

Code:

ret = avcodec_decode_video2(video_dec_ctx, frame, got_frame, &pkt);
if (ret < 0) {
    fprintf(stderr, "Error decoding video frame\n");
    return ret;
}
if (*got_frame) 
{
    printf("video_frame%s n:%d coded_n:%d pts:%s\n",
               cached ? "(cached)" : "",
               video_frame_count++, frame->coded_picture_number,
               "#"/*av_ts2timestr(frame->pts, &video_dec_ctx->time_base)*/);
    /* copy decoded frame to destination buffer:
     * this is required since rawvideo expects non aligned data */
    av_image_copy(video_dst_data, video_dst_linesize,
                  (const uint8_t **)(frame->data), frame->linesize,
                  video_dec_ctx->pix_fmt, video_dec_ctx->width, video_dec_ctx->height);
    /* write to rawvideo file */
    fwrite(video_dst_data[0], 1, video_dst_bufsize, video_dst_file);

    AVPicture pic;
    avpicture_alloc( &pic, AV_PIX_FMT_RGB24, frame->width, frame->height);
    SwsContext *ctxt = sws_getContext(frame->width, frame->height, static_cast<AVPixelFormat>(frame->format),
        frame->width, frame->height, AV_PIX_FMT_RGB24, SWS_BILINEAR, NULL, NULL, NULL);
    if ( NULL == ctxt )
    {
       //Log("failed to get sws context");
    }

    if ( 0 < sws_scale(ctxt, frame->data, frame->linesize, 0, frame->height, pic.data, pic.linesize))
    {
        char szPic[256] = { 0 };
        sprintf( szPic, "decoded/%d.bmp", video_frame_count );
        FILE *pf = fopen(szPic,"w");
        if ( NULL != pf )
        {
            BITMAPFILEHEADER bmpFileHeader = {0};
            bmpFileHeader.bfReserved1 = 0;
            bmpFileHeader.bfReserved2 = 0;
            bmpFileHeader.bfType = 0x4D42;
            bmpFileHeader.bfSize = sizeof(bmpFileHeader) + sizeof(BITMAPINFOHEADER) + pic.linesize[0] * frame->height;
            bmpFileHeader.bfOffBits = sizeof(bmpFileHeader) + sizeof(BITMAPINFOHEADER);
            BITMAPINFOHEADER bmiHeader = { 0 };
            bmiHeader.biSize = sizeof(BITMAPINFOHEADER);
            bmiHeader.biWidth = frame->width;
            bmiHeader.biHeight = 0 - frame->height;
            bmiHeader.biPlanes = 1;
            bmiHeader.biBitCount = 24;
            bmiHeader.biCompression = BI_RGB;
            bmiHeader.biSizeImage = pic.linesize[0] * frame->height;
            bmiHeader.biXPelsPerMeter = 0;
            bmiHeader.biYPelsPerMeter = 0;
            bmiHeader.biClrUsed = 0;
            bmiHeader.biClrImportant = 0;
            fwrite( &bmpFileHeader, 1, sizeof(bmpFileHeader), pf );
            fwrite( &bmiHeader, 1, sizeof(bmiHeader), pf );
            fwrite( pic.data[0], 1, pic.linesize[0] * frame->height, pf );
            fclose( pf );
        }
    }
    // pic.data[0] now contains the image data in RGB format (3 bytes)
    // and pic.linesize[0] is the pitch of the data (ie. size of a row in memory, which can be larger than width*sizeof(pixel))


    avpicture_free(&pic);
    sws_freeContext(ctxt);
}

above only decode frame then convert this from to RGB24, then write a bitmap. original video frame like this, but converted image, is there missing some code or some code is wrong?

thanks in advance.

A few things here: 1 wrong component order RGB vs BGR, 2 height in the file has to be positive, 3 line size in the file has to have minimal padding to 32-bit boundary even though SWS context stride could be larger. All together will get you correct BMP file. — Roman R.
thks, Roman R． But height can be negative, if here I give it positive value, the bitmap will be reversed image; BTW, image dimension is 1280 x 760, so it is 32-bit aligned; I also tried BGR24, same result. — Leo Hwang
The code is about right, so there is something wrong on runtime. For example, SWS context might be 1288 wide and not 1280 etc. This is OK for processing but not for the file. Perhaps you could post a link to the file you are getting. — Roman R.
the source code is from ffmpeg.org/doxygen/trunk/… width should be 1280 rather than 1288, because 1280 * 3Bytes is 4Bytes aligned — Leo Hwang

Ronald S. Bultje Ronald S. Bultje · Accepted Answer · 2015-07-22T22:15:19

fwrite( pic.data[0], 1, pic.linesize[0] * frame->height, pf );

For an image of e.g. 1280x720, linesize is typically larger, e.g. 1312, so you'll be writing more data than image size if you write linesize*height. You want to write (in a loop) width pixels offset by linesize bytes:

uint8_t *ptr = pic.data[0];
for (int y = 0; y < frame->height; y++) {
    fwrite(ptr, 1, frame->width, pf);
    ptr += pic.linesize[0];
}

And then it should work correctly.

ffmpeg sws_scale got distorted image from YUV420P to RGB24

2 Answers