A Pytorch implementation of "Cascaded Pyramid Network for Multi-Person Pose Estimation" (https://arxiv.org/abs/1711.07319).
Part of code referenced from : 'https://github.com/GengDavid/pytorch-cpn' & 'https://github.com/HRNet'
Training:
1. Download the ImageNet pre-trained Resnet weight from https://pytorch.org/docs/stable/torchvision/models.html.
and place into 'pretrain_model/pretrain_resnet/'.
2. Make sure the training & Model setting in 'experiment/experiment_0.yaml'.
3. Download the coco dataset from 'http://cocodataset.org/#home'. and place into 'dataset/'.
Install the cocoapi (https://github.com/cocodataset/cocoapi).
4. mkdir 'result/check_result',
'result/checkpoints',
'result/logs'
5. Run " training.py "
Evaluation:
0. Make sure the 'model_best.pth' is in 'result/checkpoints/'.
1. Run " evaluation.py ".
Inference:
0. Make sure the 'model_best.pth' is in 'result/checkpoints/'.
1. First we need a person detector. (maybe you can use mmdetection or detectron.)
# mmdetection: 'https://github.com/open-mmlab/mmdetection'
# detectron: 'https://github.com/facebookresearch/detectron2'
2. Put input image into 'dataset/test_data/'.
3. Run " inference.py ".