OpenPose software detects bodies in video. It runs best on GPU hardware, including on servers. An easy way to set that up is with Google’s Compute Platform, which is a metered cloud platform that can be configured with GPU hardware. There’s a couple of tricks to know, but it does give significant speed-ups over non-GPU hardware.
Launch a deep learning instance
Creating the GCP compute instance is straightforward enough, but for an easy life you want to select the specific Linux distribution: “Deep learning on Linux”. The default is (at the time of writing) Debian without any of the pre-configured GPU dependencies.
A type instance might be:
- us-central-1
- Series: N1
- GPU: V100
- Deep learning on Linux (not the default Debian); 50 GB disk
Launch that (which will cost you money), and first time you log in you’ll be promoted to install the NVIDIA drivers.
Install OpenPose
The second trick, which is specific for compatibility with OpenPose, is to remove Conda. Now, Conda is a great tool for managing Python environments but for OpenPose it includes an incompatible version of Protobuf.
To fix that, edit /etc/profile.d/env.sh and comment out:
. "/opt/conda/etc/profile.d/conda.sh"
conda activate base
…and restart your shell. For belt-and-braces I also took the nuclear option: sudo mv /opt/conda /opt/conda.deactivated.
The rest of the installation for OpenPose is as per the OpenPose docs. E.g.,
git clone https://github.com/CMU-Perceptual-Computing-Lab/openpose.git
cd openpose/
git submodule update --init --recursive --remote
sudo apt update
sudo apt-get install -y libgoogle-glog-dev libboost-all-dev libatlas-base-dev libopencv-dev protobuf-compiler
mkdir build && cd build
cmake ..
make -j`nproc`
The result
Is it worth that to be able to run on a GPU? Yes: for me, it was a 20x speed up compared to my spicy (for Intel) Mac desktop machine.
I find phrases like “20x speed up” a little too abstract to appreciate. In practice, it means something that took 30 minutes now takes 90 seconds.
