最近在研究如何快速实现图像中人体姿态的估计，也就是常见的pose estimation任务。花了些时间，实际对比了AlphaPose、BlazePose和MMPose。BlazePose主要为移动端设计，AlphaPose安装配置比较麻烦，MMPose用起来更为方便，而且支持的模型数量也很多。因此，最终选定MMPose作为人体姿态估计的算法库。以下实测环境为Ubuntu20.04系统。
MMPose源码：
https://github.com/open-mmlab/mmpose
MMPose官方文档：
https://mmpose.readthedocs.io/en/latest
(1) 准备基础的环境：
conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab
conda install pytorch torchvision -c pytorch(2) 安装mmcv，需要选择对应的cuda和torch版本，我这里是cu103和torch1.10.0：
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/index.html(3) 安装MMPose：
git clone git@github.com:open-mmlab/mmpose.git #git报错的话加sudo
cd mmpose
pip install -r requirements.txt
pip install -v -e .由于MMPose需要依赖给定的检测框，因此可以通过MMDetection先检测出图像中的人体框，再通过MMPose来输出姿态估计结果。直接采用pip安装，非常方便。
pip install mmdet以上安装完成之后，就可以在准备好的图像上测试姿态估计的效果了。这里对官方文档稍作修改，直接处理给定文件夹下的所有图像，并将结果存在另一个文件夹中。
修改之后的demo/pose_demo_mmdet.py代码如下：
# Copyright (c) OpenMMLab. All rights reserved.
import os
import warnings
from argparse import ArgumentParser
from mmpose.apis import (inference_top_down_pose_model, init_pose_model,
process_mmdet_results, vis_pose_result)
from mmpose.datasets import DatasetInfo
try:
from mmdet.apis import inference_detector, init_detector
has_mmdet = True
except (ImportError, ModuleNotFoundError):
has_mmdet = False
def main():
"""Visualize the demo images.
Using mmdet to detect the human.
"""
parser = ArgumentParser()
parser.add_argument('det_config', help='Config file for detection')
parser.add_argument('det_checkpoint', help='Checkpoint file for detection')
parser.add_argument('pose_config', help='Config file for pose')
parser.add_argument('pose_checkpoint', help='Checkpoint file for pose')
parser.add_argument('--img-root', type=str, default='', help='Image root')
parser.add_argument(
'--show',
action='store_true',
default=False,
help='whether to show img')
parser.add_argument(
'--out-img-root',
type=str,
default='',
help='root of the output img file. '
'Default not saving the visualization images.')
parser.add_argument(
'--device', default='cuda:0', help='Device used for inference')
parser.add_argument(
'--det-cat-id',
type=int,
default=1,
help='Category id for bounding box detection model')
parser.add_argument(
'--bbox-thr',
type=float,
default=0.3,
help='Bounding box score threshold')
parser.add_argument(
'--kpt-thr', type=float, default=0.3, help='Keypoint score threshold')
parser.add_argument(
'--radius',
type=int,
default=4,
help='Keypoint radius for visualization')
parser.add_argument(
'--thickness',
type=int,
default=1,
help='Link thickness for visualization')
assert has_mmdet, 'Please install mmdet to run the demo.'
args = parser.parse_args()
assert args.show or (args.out_img_root != '')
assert args.det_config is not None
assert args.det_checkpoint is not None
det_model = init_detector(
args.det_config, args.det_checkpoint, device=args.device.lower())
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
args.pose_config, args.pose_checkpoint, device=args.device.lower())
dataset = pose_model.cfg.data['test']['type']
dataset_info = pose_model.cfg.data['test'].get('dataset_info', None)
if dataset_info is None:
warnings.warn(
'Please set `dataset_info` in the config.'
'Check https://github.com/open-mmlab/mmpose/pull/663 for details.',
DeprecationWarning)
else:
dataset_info = DatasetInfo(dataset_info)
image_list = os.listdir(args.img_root)
for img_name in image_list:
image_name = os.path.join(args.img_root, img_name)
# test a single image, the resulting box is (x1, y1, x2, y2)
mmdet_results = inference_detector(det_model, image_name)
# keep the person class bounding boxes.
person_results = process_mmdet_results(mmdet_results, args.det_cat_id)
# test a single image, with a list of bboxes.
# optional
return_heatmap = False
# e.g. use ('backbone', ) to return backbone feature
output_layer_names = None
pose_results, returned_outputs = inference_top_down_pose_model(
pose_model,
image_name,
person_results,
bbox_thr=args.bbox_thr,
format='xyxy',
dataset=dataset,
dataset_info=dataset_info,
return_heatmap=return_heatmap,
outputs=output_layer_names)
if args.out_img_root == '':
out_file = None
else:
os.makedirs(args.out_img_root, exist_ok=True)
out_file = os.path.join(args.out_img_root, f'vis_{img_name}')
# show the results
vis_pose_result(
pose_model,
image_name,
pose_results,
dataset=dataset,
dataset_info=dataset_info,
kpt_score_thr=args.kpt_thr,
radius=args.radius,
thickness=args.thickness,
show=args.show,
out_file=out_file)
if __name__ == '__main__':
main()
由于测试代码参数比较长，因此新建了一个script.sh文件如下：
python demo/pose_demo_mmdet.py \
demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \
https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/hrnet_w48_coco_wholebody_384x288_dark_plus.py \
https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w48_coco_wholebody_384x288_dark-f5726563_20200918.pth \
--img-root /img_path/images/ \
--out-img-root /out_path/vis_results_whole--img-root和--out-img-root这两项分别改为原图像和保存结果的路径。
通过bash来执行测试程序：
bash script.sh最后，来看几张姿态估计的测试结果，效果还是很不错的。
