ffmpeg h265 decoder seeks past max_ts passed into avformat_seek_file

[ahmadsharif]

Summary of the bug:

I am using FFMPEG as a library. I create a AVFormatContext from a video file, and then seek to a timestamp by calling:

double timestamp;
timestamp = 0.5
avformat_seek_file(format_context, 0, INT64_MIN, timestamp, timestamp);

My understanding is that avformat_seek_file() will seek to a position where it's an I-Frame with a pts that is <= 0.5.

That is indeed what I see for most videos.

However, for certain H265 videos, FFMPEG seeks to a position beyond the max_ts timestamp passed in. To repro this behavior you can run these commands to generate a test video with just frame numbers on it:

# Create a clean conda environment for testing purposes
conda create --name test
conda activate test
conda install -c conda-forge x265

# Install some build pre-requisites
conda install pkg-config

# Build ffmpeg from source with x265 enabled
git clone https://github.com/FFmpeg/FFmpeg.git
./configure --enable-nonfree --enable-gpl --prefix=$(readlink -f ../bin)
--enable-libx265  --enable-rpath
--extra-ldflags=-Wl,-rpath=$CONDA_PREFIX/lib --enable-filter=drawtext
--enable-libfontconfig --enable-libfreetype --enable-libharfbuzz
make -j install

# Now generate a video with just frame numbers in the text per frame:
ffmpeg -f lavfi -i color=size=128x128:duration=1:rate=10:color=blue -vf
"drawtext=fontsize=30:fontcolor=white:x=(w-text_w)/2:y=(h-text_h)/2:text='Frame
%{frame_num}'" -vcodec libx265 -pix_fmt yuv420p -g 2 -crf 10 test.mp4 -y

Note that this video has 10 frames. ffprobe shows the following:

ffprobe -v error -select_streams v:0 -show_entries
frame=pts,pts_time,duration,pkt_pts_time,pkt_duration,key_frame -of csv
test.mp4
frame,1,0,0.000000,1024,1024,
frame,0,1024,0.100000,1024,1024
frame,1,2048,0.200000,1024,1024
frame,0,3072,0.300000,1024,1024
frame,1,4096,0.400000,1024,1024
frame,0,5120,0.500000,1024,1024
frame,1,6144,0.600000,1024,1024
frame,0,7168,0.700000,1024,1024
frame,1,8192,0.800000,1024,1024
frame,0,9216,0.900000,1024,1024

Now, when I open this video using FFMPEG as a library, I get an AVFormatContext. I want to decode the frame with pts=5120. So I call avformat_seek_file with min_ts=-INT64_MAX, ts=5120 and max_ts=5120.

I expect that FFMPEG will seek to the frame with pts=4096 (because it is the last keyframe before pts=5120) so I can then decode forward and eventually get frame with pts=5120 with avcodec_receive_frame(), but it seems like the first frame that I get from avcodec_receive_frame() is the one with pts=6144.

In other words:

frame,1,4096,0.400000,1024,1024 
frame,0,5120,0.500000,1024,1024 # I try to seek with max_ts=5120...
frame,1,6144,0.600000,1024,1024 # .. but FFMPEG seeks to here!

More context:
I am writing a library that wraps FFMPEG and returns frames at arbitrary timestamps. The full source code of the library is here:
​https://github.com/pytorch/torchcodec. The pull-request that reproduces
this exact scenario is here: ​https://github.com/pytorch/torchcodec/pull/178.

It would be nice if FFMPEG always seeked to a frame with pts <= the max_pts passed into avformat_seek_file. This normally does work with other codecs. Am I calling the library wrong? Should I be calling avformat_seek_file() with other flags? The documentation of avformat_seek_file is here:

https://ffmpeg.org/doxygen/7.0/group__lavf__decoding.html#ga3b40fc8d2fda6992ae6ea2567d71ba30

Here is the seek call in my code:

​https://github.com/pytorch/torchcodec/blob/dbfef1223522639d2b036a185b444eecf7748466/src/torchcodec/decoders/_core/VideoDecoder.cpp#L735

Note that I accidentally sent out this email before filing this bug. I thought I would check on the mailing list but I think a tracking bug is better for discussion as the mailing list has mostly patches.

https://ffmpeg.org/pipermail/ffmpeg-devel/2024-August/332241.html

[ahmadsharif]

The video generated by the commandline described in this bug. You can then try to seek into this file using avformat_seek_file and get the behavior I described in the bug