Adaptive graph Kolmogorov-Arnold network for 3D human pose estimation

PoseKAN, introduced on November 13, 2025, represents a breakthrough in 3D human pose estimation by overcoming the limitations of traditional graph convolutional networks (GCNs). While GCNs struggle with long-range dependencies and exhibit spectral bias, PoseKAN employs learnable functions on graph edges, allowing for adaptive feature transformations that enhance the model's expressiveness. This adaptability is vital for accurately interpreting complex pose variations, especially in challenging conditions like occlusions. The model's multi-hop feature aggregation further improves spatial awareness by enabling body joints to utilize information from both local and distant neighbors. Extensive experiments have demonstrated PoseKAN's competitive performance against state-of-the-art methods, solidifying its potential impact in the field of computer vision and human-computer interaction.