TY - JOUR
T1 - LiteSpiralGCN
T2 - Lightweight 3D hand mesh reconstruction via spiral graph convolution
AU - Wang, Yiteng
AU - Li, Minqi
AU - Zhang, Kaibing
AU - He, Xiangjian
N1 - Publisher Copyright:
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2025.
PY - 2025/5
Y1 - 2025/5
N2 - Hand mesh reconstruction technologies play an important role in computer vision, as they facilitate many applications including virtual/augmented reality, human-computer interaction, etc. However, current methods typically rely on computationally intensive architectures with excessive parameters and storage demands to achieve accuracy. In this paper, we propose a lightweight network via Spiral GCN balancing accuracy and efficiency, named LiteSpiralGCN. Our approach includes an Attention Sampling (AS) module to enhance keypoint feature interactions, a SpiralGCN module for efficient and flexible decoding, and a refinement method that leverages multi-scale and multi-stage information to boost reconstruction accuracy. Experiments conducted on benchmark datasets demonstrate that LiteSpiralGCN effectively balances parameter scale and reconstruction accuracy. Specifically, on the FreiHAND dataset, LiteSpiralGCN achieves a PA-MPJPE of 6.5 mm and a PA-MPVPE of 6.6 mm using only 9.77M parameters. Our code is publicly available at: https://github.com/minqili/LiteSpiralGCN.
AB - Hand mesh reconstruction technologies play an important role in computer vision, as they facilitate many applications including virtual/augmented reality, human-computer interaction, etc. However, current methods typically rely on computationally intensive architectures with excessive parameters and storage demands to achieve accuracy. In this paper, we propose a lightweight network via Spiral GCN balancing accuracy and efficiency, named LiteSpiralGCN. Our approach includes an Attention Sampling (AS) module to enhance keypoint feature interactions, a SpiralGCN module for efficient and flexible decoding, and a refinement method that leverages multi-scale and multi-stage information to boost reconstruction accuracy. Experiments conducted on benchmark datasets demonstrate that LiteSpiralGCN effectively balances parameter scale and reconstruction accuracy. Specifically, on the FreiHAND dataset, LiteSpiralGCN achieves a PA-MPJPE of 6.5 mm and a PA-MPVPE of 6.6 mm using only 9.77M parameters. Our code is publicly available at: https://github.com/minqili/LiteSpiralGCN.
KW - Adaptive graph convolutional networks
KW - Attention mechanism
KW - Hand mesh reconstruction
KW - Spiral convolution
UR - http://www.scopus.com/inward/record.url?scp=105003825790&partnerID=8YFLogxK
U2 - 10.1007/s10489-025-06585-0
DO - 10.1007/s10489-025-06585-0
M3 - Article
AN - SCOPUS:105003825790
SN - 0924-669X
VL - 55
JO - Applied Intelligence
JF - Applied Intelligence
IS - 7
M1 - 704
ER -