Tri-channel synergistic model for 3D human pose estimation
- 1Shandong Jianzhu University, Jinan, Shandong Province, China
- 2Shandong Huayun 3D Technology Co., Ltd., Jinan, Shandong Province, China
Abstract
3D human pose estimation from single-view images is a fundamental and challenging task in computer vision, the key of which lies in effectively capturing the complex spatial dependencies between human joint points. Existing methods usually rely on a single computational paradigm: Graph Convolutional Networks (GCN) excel at capturing local physical constraints, the self-attention mechanism (Attn) is adept at modeling global dynamic correlations, and Multi-Layer Perceptrons (MLP) provide an efficient way for global context encoding. Although previous works have attempted to fuse these methods, they often adopt serial, simple parallel or dual-stream structures, failing to fully explore the synergistic potential among these three heterogeneous computational paradigms, which results in limited representation capability under complex or rare poses. To address the above limitations, this paper proposes a novel Tri-Channel Synergistic Model (TCSM) aiming to construct a more comprehensive and robust feature representation for single-frame pose estimation. The core idea of this model is to decouple the dependencies of different natures in human poses and perform optimal parsing in independent and functionally specific channels. Specifically, our model designs three parallel feature streams: a GCN channel for encoding structure modality based on human topology, an attention channel for capturing long-range dynamic correlation modality, and an MLP channel for efficiently extracting global context modality. To achieve deep synergy, we design a two-stage feature fusion and refinement mechanism, in which the outputs of each channel interact and are iteratively enhanced. Experiments on the Human3.6M benchmark demonstrate that TCSM achieves competitive single-frame performance. The model obtains 47.7 mm MPJPE and 37.8 mm PA-MPJPE without a pose refinement module, and the ablation results verify the contribution of the tri-channel design and the two-stage interaction mechanism.
Related
- From pose estimation accuracy to motor biomarker reliability: a reliability-centred evaluation framework for markerless movement analysis
- A promptable foundation model enables automated multi-task dataset construction and real-time pose estimation in mice
- Simultaneous hyperkinetic movement disorders phenotyping: a cross-cohort pediatric transfer study using routine videos, markerless pose estimation and a tabular foundation model
- CFPR-YOLO: chili flower pose estimation for robotic pollination in unstructured environments
- Deep Learning Pose Estimation for Phenotyping of Co‐Occurring Hyperkinetic Movement Disorders
BibTeX
@article{gu2026trichannel,
title = {{Tri-channel synergistic model for 3D human pose estimation}},
author = {Ying Gu and Chenglong Li and Xinyan Gao and Jinxiao Zhang},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0359200},
url = {https://doi.org/10.1371/journal.pone.0359200},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Tri-channel synergistic model for 3D human pose estimation
AU - Ying Gu
AU - Chenglong Li
AU - Xinyan Gao
AU - Jinxiao Zhang
PY - 2026
DA - 2026/10/15
JO - PLOS One
DO - 10.1371/journal.pone.0359200
UR - https://doi.org/10.1371/journal.pone.0359200
AB - 3D human pose estimation from single-view images is a fundamental and challenging task in computer vision, the key of which lies in effectively capturing the complex spatial dependencies between human joint points. Existing methods usually rely on a single computational paradigm: Graph Convolutional Networks (GCN) excel at capturing local physical constraints, the self-attention mechanism (Attn) is adept at modeling global dynamic correlations, and Multi-Layer Perceptrons (MLP) provide an efficient way for global context encoding. Although previous works have attempted to fuse these methods, they often adopt serial, simple parallel or dual-stream structures, failing to fully explore the synergistic potential among these three heterogeneous computational paradigms, which results in limited representation capability under complex or rare poses. To address the above limitations, this paper proposes a novel Tri-Channel Synergistic Model (TCSM) aiming to construct a more comprehensive and robust feature representation for single-frame pose estimation. The core idea of this model is to decouple the dependencies of different natures in human poses and perform optimal parsing in independent and functionally specific channels. Specifically, our model designs three parallel feature streams: a GCN channel for encoding structure modality based on human topology, an attention channel for capturing long-range dynamic correlation modality, and an MLP channel for efficiently extracting global context modality. To achieve deep synergy, we design a two-stage feature fusion and refinement mechanism, in which the outputs of each channel interact and are iteratively enhanced. Experiments on the Human3.6M benchmark demonstrate that TCSM achieves competitive single-frame performance. The model obtains 47.7 mm MPJPE and 37.8 mm PA-MPJPE without a pose refinement module, and the ablation results verify the contribution of the tri-channel design and the two-stage interaction mechanism.
SN - 1932-6203
ER -