This paper presents a method for real-time multi-person 2D pose estimation using Part Affinity Fields (PAFs). The approach uses a non-parametric representation to associate body parts with individuals in an image. The architecture encodes global context and allows a greedy bottom-up parsing step that maintains high accuracy while achieving real-time performance, regardless of the number of people in the image. The architecture jointly learns part locations and their association through two branches of a sequential prediction process. The method achieved first place in the inaugural COCO 2016 keypoints challenge and significantly outperformed previous state-of-the-art results on the MPII Multi-Person benchmark in both performance and efficiency. The method takes an entire image as input and produces 2D locations of anatomical keypoints for each person. It simultaneously predicts confidence maps for body part detection and part affinity fields for part association. The confidence maps and affinity fields are parsed using a greedy inference step to output the 2D keypoints for all people in the image. The architecture consists of two branches: one predicts confidence maps, and the other predicts part affinity fields. Each branch is an iterative prediction process. The network is trained with two loss functions at each stage, one for each branch. The loss functions are weighted spatially to address the issue of incomplete labeling in some datasets. The method uses part affinity fields to encode both the location and orientation of limbs. This allows for more accurate association of body parts and eliminates false associations. The method is efficient and achieves high-quality results at a fraction of the computational cost. The method is evaluated on two benchmarks: the MPII human multi-person dataset and the COCO 2016 keypoints challenge dataset. It achieves state-of-the-art results on both benchmarks, with significant improvements in performance and efficiency compared to previous methods. The method is also efficient, achieving a speed of 8.8 fps for a video with 19 people. The method is publicly released to ensure full reproducibility and to encourage future research in the area.This paper presents a method for real-time multi-person 2D pose estimation using Part Affinity Fields (PAFs). The approach uses a non-parametric representation to associate body parts with individuals in an image. The architecture encodes global context and allows a greedy bottom-up parsing step that maintains high accuracy while achieving real-time performance, regardless of the number of people in the image. The architecture jointly learns part locations and their association through two branches of a sequential prediction process. The method achieved first place in the inaugural COCO 2016 keypoints challenge and significantly outperformed previous state-of-the-art results on the MPII Multi-Person benchmark in both performance and efficiency. The method takes an entire image as input and produces 2D locations of anatomical keypoints for each person. It simultaneously predicts confidence maps for body part detection and part affinity fields for part association. The confidence maps and affinity fields are parsed using a greedy inference step to output the 2D keypoints for all people in the image. The architecture consists of two branches: one predicts confidence maps, and the other predicts part affinity fields. Each branch is an iterative prediction process. The network is trained with two loss functions at each stage, one for each branch. The loss functions are weighted spatially to address the issue of incomplete labeling in some datasets. The method uses part affinity fields to encode both the location and orientation of limbs. This allows for more accurate association of body parts and eliminates false associations. The method is efficient and achieves high-quality results at a fraction of the computational cost. The method is evaluated on two benchmarks: the MPII human multi-person dataset and the COCO 2016 keypoints challenge dataset. It achieves state-of-the-art results on both benchmarks, with significant improvements in performance and efficiency compared to previous methods. The method is also efficient, achieving a speed of 8.8 fps for a video with 19 people. The method is publicly released to ensure full reproducibility and to encourage future research in the area.