[Paper Review] Two-person Graph Convolutional Network for Skeleton-based Human Interaction Recognition
This paper proposes a Two-person Graph Convolutional Network (2P-GCN) that models human interactions by unifying two individuals' skeletons into a single, joint graph structure, capturing both intra-body and inter-body relationships. By introducing novel edge labeling strategies and a two-person graph convolution block, 2P-GCN achieves state-of-the-art performance on four benchmarks, outperforming prior methods by up to 3.08% accuracy on NTU-120-Interaction while using only 1.47M parameters and 3.74G FLOPs.
Graph convolutional networks (GCNs) have been the predominant methods in skeleton-based human action recognition, including human-human interaction recognition. However, when dealing with interaction sequences, current GCN-based methods simply split the two-person skeleton into two discrete graphs and perform graph convolution separately as done for single-person action classification. Such operations ignore rich interactive information and hinder effective spatial inter-body relationship modeling. To overcome the above shortcoming, we introduce a novel unified two-person graph to represent inter-body and intra-body correlations between joints. Experiments show accuracy improvements in recognizing both interactions and individual actions when utilizing the proposed two-person graph topology. In addition, We design several graph labeling strategies to supervise the model to learn discriminant spatial-temporal interactive features. Finally, we propose a two-person graph convolutional network (2P-GCN). Our model achieves state-of-the-art results on four benchmarks of three interaction datasets: SBU, interaction subsets of NTU-RGB+D and NTU-RGB+D 120.
Motivation & Objective
- To address the limitation of existing GCN-based methods that treat two-person interactions as isolated single-person graphs, thereby losing critical interactive information.
- To design a unified two-person graph topology that simultaneously encodes intra-body joint connections and inter-body relationships between two individuals.
- To develop graph labeling strategies that guide the model to learn discriminative spatial-temporal interactive features.
- To propose a two-person graph convolution block (2P-GCB) that enables effective feature extraction within a single graph convolution operation.
- To achieve state-of-the-art performance on multiple benchmark datasets for skeleton-based human interaction recognition.
Proposed method
- The paper introduces a two-person graph where each frame's interaction is represented as a single graph with 2N joints (N per person), enabling joint modeling of intra- and inter-body connections.
- A relational adjacency matrix is constructed to encode both physical skeleton connections (intra-body) and inter-personal connections (inter-body), with edge weights reflecting correlation strength.
- Graph edge labeling strategies are designed to supervise the model in learning meaningful spatial-temporal interactive features, enhancing representation discrimination.
- A two-person graph convolution block (2P-GCB) is developed to perform graph convolutions on the unified two-person graph, preserving interaction-aware features across layers.
- The 2P-GCN model is evaluated on multiple GCN architectures, demonstrating consistent accuracy gains across different backbones.
- The method is scalable and generalizable, with potential extension to multi-person and human-object interaction scenarios.
Experimental results
Research questions
- RQ1Can a unified two-person graph representation improve the modeling of inter-body and intra-body relationships in skeleton-based human interaction recognition?
- RQ2How do graph labeling strategies affect the model's ability to learn discriminative spatial-temporal features for interaction recognition?
- RQ3Does replacing isolated single-person graphs with a joint two-person graph lead to consistent performance gains across different GCN architectures?
- RQ4To what extent does the proposed 2P-GCN outperform state-of-the-art methods on standard interaction recognition benchmarks?
- RQ5Can the two-person graph representation also improve performance on single-person action recognition tasks?
Key findings
- 2P-GCN achieves state-of-the-art accuracy on four benchmarks, including 97.05% on X-Sub and 98.80% on X-View of NTU-Interaction, outperforming 2S-AIGCN by 1.71% and 0.80% respectively.
- On NTU120-Interaction, 2P-GCN achieves 93.47% and 93.73% accuracy on X-Sub120 and X-Set120, surpassing 2S-AIGCN by 2.76% and 3.08%.
- The model achieves 98.90% accuracy on the SBU dataset, outperforming the previous SOTA method (AIGCN) by 0.20% and DR-GCN by 0.14%.
- Despite modeling richer interaction features, 2P-GCN maintains low complexity with only 1.47M parameters and 3.74G FLOPs, significantly lower than many SOTA models.
- The two-person graph representation improves performance not only on interaction tasks but also on single-person action recognition, demonstrating broader utility.
- The ablation study confirms that the proposed two-person graph topology consistently improves accuracy across all tested GCN models, validating its effectiveness and generalization capability.
Better researchstarts right now
From reading papers to final review, dramatically reduce your research time.
No credit card · Free plan available
This review was created by AI and reviewed by human editors.