Skip to content

Latest commit

 

History

History
61 lines (47 loc) · 2.69 KB

INSTALL.md

File metadata and controls

61 lines (47 loc) · 2.69 KB
Raw

Set-up environment

Build ffmpeg from source

Alex digged into the source code of ffmpeg and H.264 codec to support RoI encoding with an encoding quality (QP parameter in the H.264 codec) matrix file as the input. To build ffmpeg from source, please

git clone https://github.com/Alex-q-z/myh264.git

under $DIR and checkout to AccMPEG branch

git checkout AccMPEG

(Note: this branch does not support parallel encoding. Please use the master branch if you want to use RoI encoding for other purposes.)

After that, cd into the repo and run build.sh. It takes time to compile. If you compiled the code successfully, you should see something like

INSTALL libavutil/libavutil.pc

at the end.

Then, inside x264/encoder/encoder.c, search for /tank, and you'll see

    // Qizheng: add qp matrix file here
    h->operation_mode_file = x264_fopen("/tank/kuntai/code/operation_mode_file", "r");
    fscanf(h->operation_mode_file, "%d,", &h->operation_mode_flag);
    h->qp_matrix_file = x264_fopen("/tank/kuntai/code/qp_matrix_file", "r");

change the two hard-coded paths (/tank/kuntai/code/...) to $DIR/myh264/... (the path must be absolute path), and rerun build.sh (the compilation will take much less time this time, don't worry.)

Set up conda environment

First, git clone our repo under $DIR and cd into our repo, then switch to MLSys branch:

git checkout MLSys

Then, install the conda environment through the conda_env.yml:

conda env create -f conda_env.yml

Then, activate the installed environment:

conda activate accmpeg

To finish configuring the environment, you have to install pytorch and torchvision (Check https://pytorch.org/get-started/locally/ for the installation command; we used pytorch=1.8.2 with cuda 11.1 to replicate the results) and detectron2 (check https://github.com/facebookresearch/detectron2/blob/main/INSTALL.md to find a version that matches pytorch and cuda). PLease install them THROUGH pip rather than conda (since conda will install an ancient version of torchvision that triggers wierd bugs).

The procedure above will automatically install a old version ffmpeg. We need to renew it. Please download the static version of ffmpeg through https://johnvansickle.com/ffmpeg/ and use the ffmpeg binary inside to replace previous ffmpeg binary (you can check the location of it thorugh which ffmpeg.)

Then go back to our repo, cd artifact and run extract.py to extract the video dashcamcropped_1.mp4 to pngs.

Edit the config file

Then, cd .. and open settings.toml:

vim settings.toml

and edit the value of x264_dir to $DIR/myh264/