This repository includes Tensorflow code for running the multi-person pose estimiation algorithm presented in:

Alejandro Newell, Zhiao Huang, and Jia Deng, Associative Embedding: End-to-End Learning for Joint Detection and Grouping, arXiv:1611.05424, 2016.

Pretrained models are available here. Include the models in the main directory of this repository to run the demo code.

To run the code, the following must be installed:

• Python3
• tensorflow
• numpy
• Opencv3.0
• cudnn
• munkres
• tqdm
• scipy

To test a single image: python main.py --input_image_path inp.jpg --output_image_path out.jpg

For better results (but slower evaluation), you can pass --scales multi to enable multi-scale evaluation and/or -r refinement to enable an additional refinement step.

The prediction is visulized in out.jpg

To test a set of images, put your image paths in a single file, one image a line and run python main.py -l image_path_list.txt -f output.json

``output.json'' is a list of prediction per image. The data format is:

[{
     'image_path': str,
     'score': float,
     'keypoints': [x1,y1,s1,...,xk,yk,sk],
}]

Our data format is similiar to MS-COCO. Note that s_i is the confidence score of each joint instead of visibility.