CV：基于keras利用算法MobilenetV2实现局部相似域的多人二维姿态实时估计(詹姆斯扣篮+美女跳舞)

输出结果

论文复现：《Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields》

https://arxiv.org/abs/1611.08050

代码实现

更新……

import argparse

import time

import cv2

from processing import extract_parts, draw

from config_reader import config_reader

from model.cmu_model import get_testing_model

#CV：基于keras利用算法MobilenetV2实现局部相似域的多人二维姿态实时估计

#论文复现：《Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields》

if __name__ == '__main__':

   parser = argparse.ArgumentParser()

#     parser.add_argument('--image', type=str, required=True, help='input image')

   parser.add_argument('--image', type=str, default='F:/File_Python/Python_example/Human_Posture_Detection/images/ZMS03.jpg', help='input image')

   parser.add_argument('--output', type=str, default='result.png', help='output image')

   parser.add_argument('--model', type=str, default='model/keras_Realtime_Multi_Person_Pose_Estimation_model.h5', help='path to the weights file')

   args = parser.parse_args()

   image_path = args.image

   output = args.output

   keras_weights_file = args.model

   tic = time.time()

   print('start processing...')

   # load model

   # authors of original model don't use

   # vgg normalization (subtracting mean) on input images

   model = get_testing_model()

   model.load_weights(keras_weights_file)

   # load config

   params, model_params = config_reader()

   input_image = cv2.imread(image_path)  # B,G,R order

   body_parts, all_peaks, subset, candidate = extract_parts(input_image, params, model, model_params)

   canvas = draw(input_image, all_peaks, subset, candidate)

   toc = time.time()

   print('processing time is %.5f' % (toc - tic))

   cv2.imshow('keras_Realtime_Multi_Person_Pose_Estimation_model',canvas)

   cv2.waitKey()

   cv2.imwrite(output, canvas)

   cv2.destroyAllWindows()