WiCompass: Oracle-driven Data Scaling for mmWave Human Pose Estimation

Bo Liang; Chen Gong; Haobo Wang; Qirui Liu; Rungui Zhou; Fengzhi Shao; Yubo Wang; Wei Gao

cs.CVFeb 21, 2026

WiCompass: Oracle-driven Data Scaling for mmWave Human Pose Estimation

Authors: Bo Liang, Chen Gong, Haobo Wang, Qirui Liu, Rungui Zhou, Fengzhi Shao, Yubo Wang, Wei Gao,

Organizations: School of Computer Science, Peking University · University of Electronic Science and Technology of China · University of Pittsburgh · Electrical Engineering and Computer Science, Massachusetts Institute of Technology ·

Abstract

Millimeter-wave Human Pose Estimation (mmWave HPE) promises privacy but suffers from poor generalization under distribution shifts. We demonstrate that brute-force data scaling is ineffective for out-of-distribution (OOD) robustness; efficiency and coverage are the true bottlenecks. To address this, we introduce WiCompass, a coverage-aware data-collection framework. WiCompass leverages large-scale motion-capture corpora to build a universal pose space ``oracle'' that quantifies dataset redundancy and identifies underrepresented motions. Guided by this oracle, WiCompass employs a closed-loop policy to prioritize collecting informative missing samples. Experiments show that WiCompass consistently improves OOD accuracy at matched budgets and exhibits superior scaling behavior compared to conventional collection strategies. By shifting focus from brute-force scaling to coverage-aware data acquisition, this work offers a practical path toward robust mmWave sensing.

Figures & tables

Panel of Figure 1 — Figure 1. An overview of the pilot studies. (a) illustrates the relationship between model performance and size. (b) shows the results of leave-one-out generalization tests. (c) demonstrates the data efficiency of existing datasets.

Datasets	Data Format	Label Style	Size	Extra Annotations
MRI ( An et al., 2022 )	Point Cloud (PD)	COCO	160k	Subj, Scene, Action
mmBody ( Chen et al., 2022 )	PD	SMPL-X	67k	Subj, Scene
MMFi ( Yang et al., 2023 )	PD	Human3.6M	321k	Subj, Scene, Action
MilliPoint ( Cui et al., 2023 )	PD	Customized	213k	Subj, Action
HuPR ( Lee et al., 2023 )	Heatmap (HP)	LSPe	36k	No
RT-Pose ( Ho et al., 2024 )	HP & PD	Human3.6M	72k	(Multi-)Subj, Scene, Action

Table 1. Open Source mmWave HPE Datasets.

	Original paper	X-Fi ( Chen and Yang, 2024 )	mmDiff ( Fan et al., 2024 )	Our Model
mmBody ( Chen et al., 2022 )	82.00	-	68.08	67.22
MMFi ( Yang et al., 2023 )	136.1	127.4	82.73	83.42

Table 2. Comparison of MPJPE Errors Across Datasets and Models.

Figure 2. WiCompass Overview. Human poses of SMPL-X format from MoCap datasets and mmWave datasets are encoded into a shared latent space via VQ-VAE. Directional $k$ -NN coverage identifies uncovered regions, which then guides pose-conditioned data collection in real or simulated settings.

Figure 4.

k

-NN Coverage in the Latent Space.

Figure 5. Coverage-driven Data Collection Workflow.

Figure 6. Scaling Laws Comparison on Synthetic Data.

Panel of Figure 7 — Figure 7. Cross-Dataset Coverage vs. $k$ .

Panel of Figure 8 — Figure 8. NRI Distribution for Intra-Dataset Coverage Analysis.

Panel of Figure 9 — Figure 9. Real-world Validation.

Figure 10. Reconstruction Fidelity of VQ-VAE Model.

Figure 11. Pose Token Influence Visualization.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 12. Our mmWave HPE Model.

Panel of Figure 13 — Figure 13. Pose visualizations (randomly selected). From top to bottom: (1) mmBody poses, (2) poses in AMASS but uncovered by mmBody ( $k=10$ ), (3) MMFi poses, and (4) poses in AMASS but uncovered by MMFi ( $k=12$ ).

Human Pose Estimation

Human pose estimation recovers people’s body keypoints, 3D skeletons, or meshes from images, video, and other sensor data.

See all 79 papers

Expanding mmWave Datasets for Human Pose Estimation with Unlabeled Data and LiDAR Datasets
Mar 15, 2026Zhuoxuan Peng, Boan Zhu, Xingjian Zhang +2Human Pose Estimation Data Augmentation
The Hong Kong University of Science and Technology
MAEPose: Self-Supervised Spatiotemporal Learning for Human Pose Estimation on mmWave Video
Apr 30, 2026Xijia Wei, Yuan Fang, Kevin Chetty +2Masked Autoencoders Self-Supervised Learning
University College London London, UK
Wave2Body: Rethinking mmWave Human Pose Estimation as Radar-to-Body Token Translation
Jul 21, 2026Bo Liang, Chen Gong, Wei Gao +1Millimeter-Wave Radar mmWave Radar Sensing
School of Computer Science, Peking University · University of Pittsburgh · Key Laboratory of High Confidence Software Technologies, Ministry of Education (PKU)

Mar 15, 2026J/K move · Enter open · S save

Expanding mmWave Datasets for Human Pose Estimation with Unlabeled Data and LiDAR Datasets

Zhuoxuan Peng, Boan Zhu, Xingjian Zhang, Wenying Li+1

The Hong Kong University of Science and Technology

Human Pose Estimation Data Augmentation mmWave Radar Sensing

Open paper AskPDF arXiv

Current millimeter-wave (mmWave) datasets for human pose estimation (HPE) are scarce and lack diversity in both point cloud (PC) attributes and human poses, hindering the generalization ability of their trained models. On the other hand, unlabeled mmWave HPE data and diverse LiDAR HPE datasets are readily available. We propose EMDUL, a novel approach to expand the volume and diversity of an existing mmWave dataset using unlabeled mmWave data and LiDAR datasets. EMDUL consists of two independent modules, namely a pseudo-label estimator to annotate unlabeled mmWave data, and a closed-form converter that translates an annotated LiDAR PC to its mmWave counterpart. Expanding the original dataset with both LiDAR-converted and pseudo-labeled mmWave PCs significantly boosts the performance and generalization ability of all the examined HPE models, reducing 15.1% and 18.9% error for in-domain and out-of-domain settings, respectively. Code is available at https://github.com/Shimmer93/EMDUL.

Abstract

Human Pose Estimation

Explore similar work

Expanding mmWave Datasets for Human Pose Estimation with Unlabeled Data and LiDAR Datasets

MAEPose: Self-Supervised Spatiotemporal Learning for Human Pose Estimation on mmWave Video

Wave2Body: Rethinking mmWave Human Pose Estimation as Radar-to-Body Token Translation