Since r20312, FFmpeg flips jpeg if the comment in the encoded stream contains "Intel(R) JPEG Library", see attached samples inteljpeg.avi (issue 1464) and midvidjpeg.avi (issue 1468).
Attached Bad_1.jpg (Debian #625944) also contains the string, but should not be flipped.
$ ffmpeg -v 9 -loglevel 99 -i Bad_1.jpg
ffmpeg version N-32933-g9b3df9b, Copyright (c) 2000-2011 the FFmpeg developers
built on Sep 26 2011 01:22:40 with gcc 4.5.3
configuration: --cc='/usr/local/gcc-4.5.3/bin/gcc -m32'
libavutil 51. 17. 0 / 51. 17. 0
libavcodec 53. 17. 0 / 53. 17. 0
libavformat 53. 13. 0 / 53. 13. 0
libavdevice 53. 4. 0 / 53. 4. 0
libavfilter 2. 43. 5 / 2. 43. 5
libswscale 2. 1. 0 / 2. 1. 0
[mjpeg @ 0x8cbf0e0] Unsupported bit depth: 0
[mjpeg @ 0x8cbf0e0] buffer too small, expanding to 10519 bytes
[mjpeg @ 0x8cbf0e0] marker=d8 avail_size_in_buf=10519
[mjpeg @ 0x8cbf0e0] marker parser used 0 bytes (0 bits)
[mjpeg @ 0x8cbf0e0] marker=e0 avail_size_in_buf=10517
[mjpeg @ 0x8cbf0e0] marker parser used 16 bytes (128 bits)
[mjpeg @ 0x8cbf0e0] marker=fe avail_size_in_buf=10499
[mjpeg @ 0x8cbf0e0] marker parser used 46 bytes (368 bits)
[mjpeg @ 0x8cbf0e0] marker=db avail_size_in_buf=10451
[mjpeg @ 0x8cbf0e0] index=0
[mjpeg @ 0x8cbf0e0] qscale[0]: 3
[mjpeg @ 0x8cbf0e0] marker parser used 67 bytes (536 bits)
[mjpeg @ 0x8cbf0e0] marker=db avail_size_in_buf=10382
[mjpeg @ 0x8cbf0e0] index=1
[mjpeg @ 0x8cbf0e0] qscale[1]: 6
[mjpeg @ 0x8cbf0e0] marker parser used 67 bytes (536 bits)
[mjpeg @ 0x8cbf0e0] marker=c4 avail_size_in_buf=10313
[mjpeg @ 0x8cbf0e0] class=0 index=0 nb_codes=12
[mjpeg @ 0x8cbf0e0] class=1 index=0 nb_codes=251
[mjpeg @ 0x8cbf0e0] class=0 index=1 nb_codes=12
[mjpeg @ 0x8cbf0e0] class=1 index=1 nb_codes=251
[mjpeg @ 0x8cbf0e0] marker parser used 418 bytes (3344 bits)
[mjpeg @ 0x8cbf0e0] marker=c0 avail_size_in_buf=9893
[mjpeg @ 0x8cbf0e0] sof0: picture: 300x300
[mjpeg @ 0x8cbf0e0] component 0 2:2 id: 0 quant:0
[mjpeg @ 0x8cbf0e0] component 1 1:1 id: 1 quant:1
[mjpeg @ 0x8cbf0e0] component 2 1:1 id: 2 quant:1
[mjpeg @ 0x8cbf0e0] pix fmt id 22111100
[mjpeg @ 0x8cbf0e0] marker parser used 17 bytes (136 bits)
[mjpeg @ 0x8cbf0e0] escaping removed 255 bytes
[mjpeg @ 0x8cbf0e0] marker=da avail_size_in_buf=9874
[mjpeg @ 0x8cbf0e0] component: 0
[mjpeg @ 0x8cbf0e0] component: 1
[mjpeg @ 0x8cbf0e0] component: 2
[mjpeg @ 0x8cbf0e0] marker parser used 9618 bytes (76938 bits)
[mjpeg @ 0x8cbf0e0] marker=d9 avail_size_in_buf=0
[mjpeg @ 0x8cbf0e0] mjpeg decode frame unused 0 bytes
Input #0, image2, from 'Bad_1.jpg':
Duration: 00:00:00.04, start: 0.000000, bitrate: N/A
Stream #0:0, 1, 1/25: Video: mjpeg, yuvj420p, 300x300 [SAR 1:1 DAR 1:1], 1/25, 25 tbr, 25 tbn, 25 tbc
At least one output file must be specified
In this case the file is flipped unnecessary.
libavcodec/mjpegdec.c:
else if((len > 20 && !strncmp(cbuf, "Intel(R) JPEG Library", 21)) || (len > 19 && !strncmp(cbuf, "Metasoft MJPEG Codec", 20))){ - s->flipped = 1; + s->flipped = 0; }so it looks like the intel jpeg encoder can save flipped and not flipped jpeg frames and the question is how to identify which frames should be fliped and which not.