stefanopini / simple-HRNet

Multi-person Human Pose Estimation with HRNet in Pytorch

Multi-person Human Pose Estimation with HRNet in PyTorch

This is an unofficial implementation of the paper Deep High-Resolution Representation Learning for Human Pose Estimation.
The code is a simplified version of the official code with the ease-of-use in mind.

The code is fully compatible with the official pre-trained weights and the results are the same of the original implementation (only slight differences on gpu due to CUDA). It supports both Windows and Linux.

This repository provides:

• A simple HRNet implementation in PyTorch (>=1.0) - compatible with official weights (pose_hrnet_*).
• A simple class (SimpleHRNet) that loads the HRNet network for the human pose estimation, loads the pre-trained weights, and make human predictions on a single image or a batch of images.
• Support for "SimpleBaselines" model based on ResNet - compatible with official weights (pose_resnet_*).
• Support for multi-GPU inference.
• Add options for retrieving yolo bounding boxes and HRNet heatmaps.
• NEW Multi-person support with YOLOv3 (enabled by default), YOLOv3-tiny, or YOLOv5 by Ultralytics.
• A reference code that runs a live demo reading frames from a webcam or a video file.
• A relatively-simple code for training and testing the HRNet network.
• A specific script for training the network on the COCO dataset.
• NEW An updated Jupyter Notebook compatible with Google Colab showcasing how to use this repository.
  • Click here to open the notebook on Colab!
  • Thanks to @basicvisual and @wuyenlin for the initial notebook.
• NEW Support for TensorRT (thanks to @gpastal24, see #99 and #100).

If you are interested in HigherHRNet, please look at simple-HigherHRNet

Examples

Class usage

import cv2
from SimpleHRNet import SimpleHRNet

model = SimpleHRNet(48, 17, "./weights/pose_hrnet_w48_384x288.pth")
image = cv2.imread("image.png", cv2.IMREAD_COLOR)

joints = model.predict(image)

The most useful parameters of the __init__ function are:

c	number of channels (HRNet: 32, 48; PoseResNet: resnet size)
nof_joints	number of joints (COCO: 17, MPII: 16)
checkpoint_path	path of the (official) weights to be loaded
model_name	'HRNet' or 'PoseResNet'
resolution	image resolution, it depends on the loaded weights
multiperson	enable multiperson prediction
return_heatmaps	the `predict` method returns also the heatmaps
return_bounding_boxes	the `predict` method returns also the bounding boxes (useful in conjunction with `multiperson`)
max_batch_size	maximum batch size used in hrnet inference
device	device (cpu or cuda)

Running the live demo

From a connected camera:

python scripts/live-demo.py --camera_id 0

From a saved video:

python scripts/live-demo.py --filename video.mp4

For help:

python scripts/live-demo.py --help

Extracting keypoints:

From a saved video:

python scripts/extract-keypoints.py --format csv --filename video.mp4

For help:

python scripts/extract-keypoints.py --help

Converting the model to TensorRT:

Warning: require the installation of TensorRT (see Nvidia website) and onnx. On some platforms, they can be installed with

pip install tensorrt onnx

Converting in FP16:

python scripts/export-tensorrt-model.py --device 0 --half

For help:

python scripts/export-tensorrt-model.py --help

Running the training script

python scripts/train_coco.py

For help:

python scripts/train_coco.py --help

Installation instructions

• Clone the repository
  git clone https://github.com/stefanopini/simple-HRNet.git

• Install the required packages
  pip install -r requirements.txt

• Download the official pre-trained weights from https://github.com/leoxiaobin/deep-high-resolution-net.pytorch
  Direct links (official Drive folder, official OneDrive folder):

  • COCO w48 384x288 (more accurate, but slower) - Used as default in live_demo.py and the other scripts
    pose_hrnet_w48_384x288.pth
  • COCO w32 256x192 (less accurate, but faster)
    pose_hrnet_w32_256x192.pth
  • MPII w32 256x256 (MPII human joints)
    pose_hrnet_w32_256x256.pth

  Remember to set the parameters of SimpleHRNet accordingly (in particular c, nof_joints, and resolution).

• For multi-person support:

  • Get YOLOv3:
    • Clone YOLOv3 in the folder ./models/detectors and change the folder name from PyTorch-YOLOv3 to yolo
      OR
    • Update git submodules
      git submodule update --init --recursive
  • Install YOLOv3 required packages
    pip install -r requirements.txt (from folder ./models/detectors/yolo)
  • Download the pre-trained weights running the script download_weights.sh from the weights folder

• (Optional) Download the COCO dataset and save it in ./datasets/COCO

• Your folders should look like:

  simple-HRNet
  ├── datasets                (datasets - for training only)
  │  └── COCO                 (COCO dataset)
  ├── losses                  (loss functions)
  ├── misc                    (misc)
  │  └── nms                  (CUDA nms module - for training only)
  ├── models                  (pytorch models)
  │  └── detectors            (people detectors)
  │    └── yolo               (PyTorch-YOLOv3 repository)
  │      ├── ...
  │      └── weights          (YOLOv3 weights)
  ├── scripts                 (scripts)
  ├── testing                 (testing code)
  ├── training                (training code)
  └── weights                 (HRnet weights)
  

• If you want to run the training script on COCO scripts/train_coco.py, you have to build the nms module first.
  Please note that a linux machine with CUDA is currently required. Build it with either:

  • cd misc; make or
  • cd misc/nms; python setup_linux.py build_ext --inplace

  You may need to add the ./misc/nms directory in the PYTHONPATH variable:
  export PYTHONPATH="<path-to-simple-HRNet>/misc/nms:$PYTHONPATH"

Google Colab notebook

Thanks to the great work of @basicvisual and @wuyenlin, you can also try this repository online on Google Colab. More details and the notebook URL are available in this issue.
Please make sure to make a copy on your own Google Drive and to change the Colab "Runtime type" from CPU to GPU or TPU.