Manuscript received July 9, 2024; revised August 12, 2024; accepted September 26, 2024; published March 14, 2025.

Abstract—Human pose estimation is an essential component of computer vision systems involving human activity, as it is concerned with predicting the configuration of the human body in 2D or 3D coordinates. The pose is expressed as related keypoints representing body parts. We consider the case of 2D bottom-up pose estimation, where the location of identity-free keypoints is predicted and then grouped into individual persons. In contemporary work, while the keypoint prediction process is learnable and automated, keypoint-grouping still largely relies on non-learnable optimization algorithms operating in embedding space, grouping similar keypoints regardless of the resulting pose structure. To overcome this limitation, this paper presents the Graph Edge Classifier (GEC), a novel, learnable keypoint-grouping method. In GEC, predicted keypoints are represented as a graph, where each keypoint is connected to all potentially related keypoints via edges. The main objective is to classify edges as either connected or not connected. The method consists of three components: A Graph Neural Network (GNN) encoder for node and edge feature learning, an edge classifier network (decoder), and a post-processing step. Additionally, we introduce two novel update functions for node and edge features within the message-passing neural network framework. Our method achieves an Average Precision (AP) score of 46.1% on the CrowdPose test set, which is comparable to similar bottom-up methods. Moreover, our model is lightweight, with only 0.274 million parameters, making it more efficient in terms of computational resources compared to other learnable keypoint-grouping methods. The learnable, efficient, and structure-aware nature of our approach offers potential for further improvement, especially through integrated end-toend training of both the backbone and grouping networks.

Keywords—2D human pose estimation, keypoint grouping, message-passing neural network, graph neural networks, edge classification

Cite: Marwan Maher, Radwa Fathalla, and Mohamed Shaheen, "Graph Edge Classification for Keypoint Grouping in Multi-person Pose Estimation," Journal of Image and Graphics, Vol. 13, No. 2, pp. 130-139, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.