Batch processing of images for pose estimation via OpenPose (currently only body keypoints). Implemented in PyTorch.
Based on https://github.com/Hzzone/pytorch-openpose, but with simplification, support for batch processing and conda-based setup.
To create a compatible anaconda environment, clone this repo and run bash setup_openpose_env.sh in the directory.
Alternatively, its contents can be run line-wise at the command line:
conda create -n openpose_pytorch -y python=3.8
source ~/anaconda3/etc/profile.d/conda.sh
conda activate openpose_pytorch
conda install -y pytorch torchvision cudatoolkit=10.2 -c pytorch
conda install -y ipykernel
conda install -y matplotlib
pip install wget
Activate the environment if your not already in it (conda activate openpose_pytorch) and run the setup script (python3 setup.py install).
Note: has not been tested in Windows or macOS.
import numpy as np
import torch
from openpose_pytorch.models import body_pose
from openpose_pytorch.utils import load_and_transform, upscale_unpad
from openpose_pytorch.keypoints import get_candidates_subsets, get_keypoints
import glob
from matplotlib import pyplot as plt
model = body_pose(pretrained=True)
images_list = sorted(glob.glob('/path/to/images/directory/*'))
images = load_and_transform(images_list, resolution=(250,250))
if torch.cuda.is_available():
model = model.cuda()
images = images.cuda()
pafs, heatmaps = model(images)
orig_size = plt.imread(images_list[0]).shape[:2]
pafs = upscale_unpad(pafs, orig_size)
heatmaps = upscale_unpad(heatmaps, orig_size)
candidates, subsets = get_candidates_subsets(heatmaps, pafs)
keypoints = get_keypoints(candidates, subsets)
This will return a list of n_people x 18 x 3 keypoints tensors for each input images.
The majority of computational time is in the get_candidates_subsets step. Refactoring could substantially improve runtime.
face and hand keypoints
higher-level wrapper
detection of isolated parts (e.g. hands + arm)
segmentation approximation