Lite-HRNet is a 2D pose estimation method whose implementation depends on several packages from OpenMMLab, notably MMCV and MMPose. Since Lite-HRNet was made publicly available, installation of the method has become more complicated due to major upgrades to MMCV and MMPose. In addition, by default Lite-HRNet is compatible with the COCO and MPII Human Pose datasets but not with the SPEED+ dataset. The method must therefore be extended to make it compatible with the SPEED+ dataset.
conda create -n openmmlab python=3.8
conda activate openmmlab
Install PyTorch 1.11.0 using following command:
conda install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch
For more information, see the installation instructions on the PyTorch website.
Install MIM, a package management software, using the following command:
pip install -U openmim
Install version 1.7.0 of the old release of MMCV using the following command:
mim install mmcv-full==1.7.0
For more information, see the installation instructions here.
Install version 0.29.0 of the old release of MMPose using the following command:
pip install mmpose==0.29.0
For more information, see the installation instructions here.
git clone https://github.com/HRNet/Lite-HRNet.git
pip install -r requirements.txt
For more information, see the instructions here.
-
Add the
topdown_speedplus_dataset.py
file from this GitHub repository to themmpose/datasets/datasets/top_down
folder -
When running Lite-HRNet for the first time, modify line 15 of the
topdown_speedplus_dataset.py
file in the following way:
@DATASETS.register_module(name='TopDownSpeedPlusDataset')
- Update the
__init__.py
file in themmpose/datasets/datasets/top_down
andmmpose/datasets/datasets
folders
For more information, see the instructions here.
-
Add the
my_transforms.py
file from this GitHub repository to themmpose/datasets/pipelines
folder -
When running Lite-HRNet for the first time, modify line 13 of the
my_transforms.py
file in the following way:
@PIPELINES.register_module(name='TopDownBBoxCrop')
- Update the
__init__.py
file in themmpose/datasets/pipelines
folder
For code modifications required because of MMCV and MMPose upgrades, see the following files in this GitHub repository:
- configs/base/datasets/speedplus.py
- models/backbones/litehrnet.py
- tools/train.py
- tools/test.py