Organizations: School of Computer Science, Northwestern Polytechnical University · Xidian University · School of Cybersecurity, Northwestern Polytechnical University
State-of-the-art whole-body pose estimators often lack robustness, producing anatomically implausible predictions in challenging scenes. We posit this failure stems from spurious correlations learned from visual context, a problem we formalize using a Structural Causal Model (SCM). The SCM identifies visual context as a confounder that creates a non-causal backdoor path, corrupting the model's reasoning. We introduce the Causal Intervention Graph Pose (CIGPose) framework to address this by approximating the true causal effect between visual evidence and pose. The core of CIGPose is a novel Causal Intervention Module: it first identifies confounded keypoint representations via predictive uncertainty and then replaces them with learned, context-invariant canonical embeddings. These deconfounded embeddings are processed by a hierarchical graph neural network that reasons over the human skeleton at both local and global semantic levels to enforce anatomical plausibility. Extensive experiments show CIGPose achieves a new state-of-the-art on COCO-WholeBody. Notably, our CIGPose-x model achieves 67.0% AP, surpassing prior methods that rely on extra training data. With the additional UBody dataset, CIGPose-x is further boosted to 67.5% AP, demonstrating superior robustness and data efficiency. The codes and models are publicly available at https://github.com/53mins/CIGPose.
Figures & tables
Figure 1: (a) Comparison of CIGPose with related models for whole-body pose estimation on COCO-WholeBody. (b) Qualitative comparison between CIGPose-x and RTMPose-x [ 16 ] .
Figure 2: (a) The proposed Structural Causal Model (SCM) for keypoint estimation, (b) The intervened SCM after applying the do -operator to the keypoint embeddings, (c) The realization of each component within our Causal Intervention Module (CIM).
Figure 3: Validation of confounder score sc(k) on COCO-WholeBody [ 19 ] . The confounder score is significantly higher for occluded keypoints than for visible ones, for both (a) easily occluded limb keypoints and (b) all keypoints combined.
Figure 4: UMAP visualization of initial contextual embeddings vs. their corresponding learned canonical embeddings.
Figure 5: Overview of our CIGPose architecture. During training, embeddings are processed in two paths: (1) a counterfactual path where our CIM module deconfounds them before the Hierarchical GNN, and (2) an observational path using original embeddings for consistency. Inference relies solely on the counterfactual path.
Method
Input Size
GFLOPs
whole-body
body
foot
face
hand
AP
AR
AP
AR
AP
AR
AP
AR
AP
AR
Whole-
SN† [ 12 ]
N/A
272.3
32.7
45.6
42.7
58.3
9.9
36.9
64.9
69.7
40.8
58.0
body
OpenPose [ 3 ]
N/A
451.1
44.2
52.3
56.3
61.2
53.2
64.5
76.5
84.0
38.6
43.3
Bottom-
PAF† [ 2 ]
512 × 512
329.1
29.5
40.5
38.1
52.6
5.3
27.8
65.6
70.1
35.9
52.8
up
AE [ 32 ]
512 × 512
212.4
44.0
54.5
58.0
66.1
57.7
72.5
58.8
65.4
48.1
57.4
SimpleBaseline [ 50 ]
384 × 288
20.4
57.3
67.1
66.6
74.7
63.5
76.3
73.2
81.2
53.7
64.7
Table 1: Whole-body pose estimation results on COCO-WholeBody [ 19 , 51 ] V1.0 dataset. “*” denotes the model that relies on two-stage distillation and additional training data from the UBody dataset [ 27 ] . “†” indicates multi-scale testing. Flip test is used.
Method
#Params
GFLOPs
COCO Val
AP
AR
SimpleBaseline [ 50 ]
60M
15.7
73.5
79.0
HRNet† [ 43 ]
64M
32.9
76.3
81.2
HRFormer-B† [ 59 ]
43M
26.8
77.2
82.0
ViTPose++-B [ 53 ]
86M
17.1
77.0
82.6
RTMPose-m [ 16 ]
14M
1.9
75.8
80.6
Table 2: Comparisons of CIGPose and SOTA methods on COCO val set [ 28 ] . The default input resolution is 256 × 192, “†” denotes the input resolution is 384 × 288.
Method
#Params
AP
AP E
AP M
AP H
Bottom-up methods
OpenPose [ 3 ]
-
-
62.7
48.7
32.3
HrHRNet [ 4 ]
63.8M
65.9
73.3
66.5
57.9
DEKR [ 10 ]
65.7M
67.3
74.6
68.1
58.7
SWAHR [ 31 ]
63.8M
71.6
78.9
72.4
63.0
Top-down methods
Table 3: Comparisons with SOTA methods on the CrowdPose [ 24 ] dataset. The default input resolution is 256 × 192, “†” denotes the input resolution is 384 × 288.
Figure 6: Qualitative comparison of CIGPose-x against the baseline RTMPose-x [ 16 ] on challenging images. From left to right: input image, RTMPose-x, and CIGPose-x.
CIM
Gh
Gp
Model
AP
AR
✓
✓
✓
CIGPose-l
66.3 (+1.5)
74.9 (+1.9)
CIGPose-x
67.0 (+1.7)
75.4 (+2.1)
✗
✓
✓
CIGPose-l
66.0 (+1.2)
74.8 (+1.8)
CIGPose-x
66.8 (+1.5)
75.5 (+2.2)
✗
✗
✓
CIGPose-l
65.6 (+0.8)
74.6 (+1.6)
CIGPose-x
66.3 (+1.0)
75.0 (+1.7)
Table 4: Ablation study of CIGPose on COCO-WholeBody. Values in parentheses are AP/AR gains.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Easy-drop p
Kept inst.
Mean Δˉ (px)
95% CI
0.00
104,125
8.47
[8.20, 8.72]
0.30
72,888
10.14
[9.78, 10.48]
0.50
52,063
10.45
[10.02, 10.87]
Appendix
Table 5: Within instance top- n enrichment on COCO-WholeBody. Keypoints selected by sc exhibit larger localization error than the remaining keypoints in the same instance. Easy-drop p denotes the fraction of easiest instances removed when forming harder subsets, and 95% CIs are percentile bootstrap intervals over instances.
Figure 7: Visualization of posterior probability distributions. (A) For an unconfounded keypoint (green circle, nose), the distributions exhibit sharp, high peaks, indicating low uncertainty. (B) For a confounded keypoint (red circle, left ankle), which is occluded and in shadow, the distributions are diffuse with low peaks, signaling high predictive ambiguity.
Hyperparameter
Value
Optimizer
AdamW
Base Learning Rate
2×10−3
Weight Decay
0.05
LR Schedule
Cosine Annealing
Warm-up Iterations
1000
Max Gradient Norm
35
Appendix
Table 6: Key training hyperparameters for the CIGPose-x model.
Body Part
Intervention Freq. (%)
Face
0.67
Torso
0.16
Arms
0.24
Hands
0.90
Legs
0.89
Feet
1.36
Appendix
Table 7: Intervention frequency per body part on COCO-WholeBody. The module intervenes most often on extremities, which are most prone to confounding.
λ
Whole-Body AP (%)
0
65.8
0.01
66.1
0.1
66.3
0.5
65.9
Appendix
Table 8: Effect of the counterfactual consistency loss weight λ .
Intervention Strategy (Training)
Whole-Body AP (%)
Threshold τ=0.7
65.9
Threshold τ=0.8
65.8
Top- n ( n=11 )
66.1
Top- n ( n=13 )
66.3
Top- n ( n=15 )
66.0
Appendix
Table 9: Comparison of intervention strategies during training. The fixed-budget ‘top- n ‘ strategy provides a more stable and effective training signal, with n=13 yielding the best result on COCO-WholeBody.