DSR-Net: Learning Dynamic Skeletal Relations for 3D Human Pose Estimation

Fan Li; Jian Xiang; LinLin Xue; Yuwei Gao

cs.CVOct 10, 2026

DSR-Net: Learning Dynamic Skeletal Relations for 3D Human Pose Estimation

Authors: Fan Li, Jian Xiang, LinLin Xue, Yuwei Gao

Organizations: School of Computer Science and Technology, Zhejiang University of Science and Technology, Liuxia, Hangzhou 310023, China

Abstract

3D human pose estimation requires accurate modeling of motion-dependent skeletal structures and inter-joint dependencies. Existing methods typically rely on fixed skeletal topologies in Graph Convolutional Networks (GCNs) or implicitly capture joint relations through attention mechanisms, limiting their adaptability to diverse motion patterns. To address this issue,we proposeDSR-Net, a dynamic structural representation learning framework for 3D human pose estimation. Specifically, a Dynamic Skeletal Relation (DSR) module is introduced to learn adaptive inter-joint dependencies beyond predefined skeletal connections. Based on the learned relations, a hierarchical relation-aware fusion strategy is developed to facilitate effective interaction between structural and temporal representations. In addition, a lightweight Residual Skeletal Recalibration Module (RSRM) is designed to enhance the stability of structure-aware features with minimal computational overhead. Extensive experiments on Human3.6M and MPI-INF-3DHP demonstrate that DSR-Net achieves MPJPEs of 37.9 mm and 14.9 mm, respectively, consistently improving over strong baselines. The results validate the effectiveness of dynamic skeletal relation modeling for robust 3D human pose estimation.

Figures & tables

Figure 1: Overall architecture of the proposed framework.

Figure 2: Overview of the proposed GCN-DSR module. The framework first constructs sparse dynamic graph topologies through dynamic relation learning, and then performs graph convolution-based feature propagation in a decoupled manner. The same module is shared in both spatial and temporal modes.

Figure 3: Architecture of the proposed RSRM module.

Figure 4: Comparison with representative 3D human pose estimation methods on the Human3.6M dataset. MPJPE denotes the mean per-joint position error (lower is better), while Param indicates the number of model parameters.

Method	T	Param (M)	MACs (G)	P1 $\downarrow$	P2 $\downarrow$
MHFormer [ 11 ]	351	30.9	7.0	43.0	34.4
MixSTE [ 29 ]	243	33.6	139.0	40.9	32.6
STCFormer [ 23 ]	243	4.7	19.6	41.0	32.0
STCFormer-L [ 23 ]	243	18.9	78.2	40.5	31.8
PoseFormerV2 [ 33 ]	243	14.3	0.5	45.2	35.6
GLA-GCN [ 28 ]	243	1.3	1.5	44.4	34.8

Table 1: Quantitative comparison results on Human3.6M. T: Number of input frames; P1: MPJPE error (mm); P2: P-MPJPE error (mm); Param and MACs indicate the number of parameters and computational complexity, respectively. The best and second-best results are shown in bold and underlined, respectively.

Method	T	Dir.	Disc.	Eat.	Greet	Phone	Photo	Pose	Pur.	Sit	SitD.	Smoke	Wait	WalkD.	Walk	WalkT.	Avg.
MPJPE (P1)
SemGCN [ 32 ]	1	48.2	60.8	51.8	64.0	64.6	53.6	51.1	67.4	88.7	57.7	73.2	65.6	48.9	64.8	51.9	60.8
EHGFormer [ 34 ]	243	34.1	38.0	38.2	32.4	40.6	47.8	37.3	34.3	50.2	56.1	39.4	36.7	26.3	35.8	25.8	38.2
MotionBERT [ 35 ]	243	36.3	38.7	38.6	33.6	42.1	50.1	36.2	35.7	50.1	56.6	41.3	37.4	37.7	25.6	26.5	39.2
MotionAGFormer [ 18 ]	243	36.4	38.4	36.8	32.9	40.9	48.5	36.6	34.6	51.7	52.8	41.0	36.4	36.5	26.7	27.0	38.4
DSR-Net (Ours)	243	36.0	38.4	36.9	33.2	40.7	47.2	36.2	34.6	48.7	51.5	40.5	36.7	36.0	25.5	25.9	37.9

Table 2: MPJPE results (mm) on the Human3.6M dataset under Protocol #1 (P1) and Protocol #2 (P2). All methods use SH-detected 2D keypoints for a fair comparison.

Method	T	PCK $\uparrow$	AUC $\uparrow$	MPJPE $\downarrow$
STCFormer [ 23 ]	81	98.7	83.9	23.1
PoseFormerV2 [ 33 ]	81	97.9	78.8	27.8
GLA-GCN [ 28 ]	81	98.5	79.1	27.7
DiffPose [ 4 ]	81	98.0	75.9	29.1
CGFusionFormer [ 16 ]	81	97.9	78.5	27.2
STAPFormer [ 31 ]	81	98.8	83.9	20.5

Table 3: Quantitative comparison with existing methods on the MPI-INF-3DHP dataset. PCK and AUC denote Percentage of Correct Keypoints and Area Under the Curve, respectively. “T” represents the number of input frames.

GCN	RSRM	HRF	MPJPE	$\Delta$
$\times$	$\times$	$\times$	39.16	–
$\checkmark$	$\times$	$\times$	38.54	0.62
$\checkmark$	$\checkmark$	$\times$	38.24	0.92
$\checkmark$	$\checkmark$	$\checkmark$	37.90	1.26

Table 4: Progressive integration study of the proposed modules.

Figure 5: Ablation study of different components on the Human3.6M dataset. Each proposed module progressively improves the pose estimation performance, leading to a cumulative reduction of 1.26 mm in MPJPE.

Dynamic Spatial Relation	Temporal Constraint	Graph Structure	MPJPE
$\times$	$\times$	Static Graph	39.16
$\checkmark$	$\times$	Dynamic Spatial Graph	39.52
$\checkmark$	$\checkmark$	Dynamic Spatiotemporal Graph	38.54

Table 5: Analysis of different dynamic relation modeling configurations.

Recalibration Strategy	Insertion Position	MPJPE
SE	Fusion Head	39.16
RSRM	Fusion Head	38.85
RSRM	Intra-Block	38.61

Table 6: Analysis of different recalibration strategies and insertion positions.

Intra-Block Fusion	Cross-Stream Fusion	MPJPE
Weighted Summation	Weighted Summation	39.16
Weighted Summation	Dynamic Channel Fusion	38.44
Learnable Gating	Dynamic Channel Fusion	38.00

Table 7: Analysis of hierarchical fusion mechanisms.

Figure 6: Spatial visualization of learned dynamic skeletal relations (Top-K).

Figure 7: Visualization of the learned skeletal relations at different network layers (Layer 3, Layer 6, and Layer 11).

Figure 8: Visualization of learned dynamic skeletal relations under different actions: (a) Walk, (b) Eating, and (c) Sitting.

Figure 9: Qualitative comparison of 3D pose estimation results between the proposed method and MixTGFormer on the Human3.6M dataset.

Figure 10: Qualitative results of DSR-Net on the 3DPW dataset under challenging in-the-wild scenarios, including outdoor environments, motion variations, and partial occlusions.

Explore similar work

Motion-Adaptive Multi-Scale Temporal Modelling with Skeleton-Constrained Spatial Graphs for Efficient 3D Human Pose Estimation
Apr 4, 2026Ruochen Li, Shuang Chen, Wenke E +23D Human Pose Estimation Graph Neural Networks
Department of Computer Science, Durham University, UK
Dual-stream Spatio-Temporal GCN-Transformer Network for 3D Human Pose Estimation
Apr 20, 2026Jiawen Duan, Jian Xiang, Zhiqiang Li +23D ViTs 3D Reconstruction
School of Information and Electronic Engineering, Zhejiang University of Science and Technology, Liuxia, Hangzhou 310023, China
PRGCN: A Graph Memory Network for Cross-Sequence Pattern Reuse in 3D Human Pose Estimation
Oct 22, 2025Zhuoyang Xie, Yibo Zhao, Hui Huang +2Memory-Augmented Neural Networks Graph Neural Networks
College of Computer Science and Artificial Intelligence, Wenzhou University, Wenzhou, 325035, P.R China. · the Key Laboratory of Computer Vision and System, Ministry of Education, Tianjin University of Technology, Tianjin, 300384, China. · School of Data Science and Artificial Intelligence, Wenzhou University of Technology, Wenzhou, 325000, P.R China. +1

Apr 4, 2026J/K move · Enter open · S save

Motion-Adaptive Multi-Scale Temporal Modelling with Skeleton-Constrained Spatial Graphs for Efficient 3D Human Pose Estimation

Ruochen Li, Shuang Chen, Wenke E, Farshad Arvin+1

Department of Computer Science, Durham University, UK

3D Human Pose Estimation Graph Neural Networks

Open paper AskPDF arXiv

Accurate 3D human pose estimation from monocular videos requires effective modelling of complex spatial and temporal dependencies. However, existing methods often face challenges in efficiency and adaptability when modelling spatial and temporal dependencies, particularly under dense attention or fixed modelling schemes. In this work, we propose MASC-Pose, a Motion-Adaptive multi-scale temporal modelling framework with Skeleton-Constrained spatial graphs for efficient 3D human pose estimation. Specifically, it introduces an Adaptive Multi-scale Temporal Modelling (AMTM) module to adaptively capture heterogeneous motion dynamics at different temporal scales, together with a Skeleton-constrained Adaptive GCN (SAGCN) for joint-specific spatial interaction modelling. By jointly enabling adaptive temporal reasoning and efficient spatial aggregation, our method achieves strong accuracy with high computational efficiency. Extensive experiments on Human3.6M and MPI-INF-3DHP datasets demonstrate the effectiveness of our approach.

Abstract

Explore similar work

Motion-Adaptive Multi-Scale Temporal Modelling with Skeleton-Constrained Spatial Graphs for Efficient 3D Human Pose Estimation

Dual-stream Spatio-Temporal GCN-Transformer Network for 3D Human Pose Estimation

PRGCN: A Graph Memory Network for Cross-Sequence Pattern Reuse in 3D Human Pose Estimation