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.
Introduction
Millimeter-wave (mmWave) radar is a compelling alternative to cameras for human pose estimation, offering privacy protection and functionality in low-light conditions. However, current models are often "brittle," failing to generalize when users perform actions slightly different from those in the training set. While the standard response is to "just collect more data," this research shows that the bottleneck is not data volume, but rather efficiency and coverage.
The Data Bottleneck: Redundancy and Capacity
The authors identify two major misconceptions in current mmWave HPE research:
- Model Size: Increasing the parameters of a Point Transformer model beyond ~10MB yields diminishing returns. The system's intelligence is limited by data diversity, not neural network capacity.
- The 70% Redundancy Rule: In a progressive subset study, researchers found they could discard 70% of training data with less than a 2% drop in performance. This confirms that existing datasets are heavily bloated with redundant information.
Generalization Challenges
The primary hurdle for robust sensing is the generalization gap. Models perform well on seen actions (In-Distribution) but fail on unseen ones (Out-of-Distribution). This "symmetry blindness"—where a model might recognize a left-hand wave but not a right-hand one—highlights the need for a more structured approach to data diversity.
WiCompass: A Coverage-Aware Framework
WiCompass is a closed-loop framework designed to navigate the "manifold" of human motion:
- Universal Pose Tokenizer: A VQ-VAE trained on the AMASS dataset (8.9M frames) creates a modality-agnostic "vocabulary" of human poses.
- Coverage Analysis: By projecting radar data into this latent space, WiCompass identifies "gaps" in the universal pose space.
- Normalized Redundancy Index (NRI): A new metric to quantify dataset repetition.
- Guided Acquisition: Using Capped-PPS sampling, the framework prioritizes collecting specific "gap" poses, realized through simulators or human volunteers.
Key Results
- Superior Scaling: WiCompass exhibited a stronger scaling exponent ($\approx D^{-0.0116}$) compared to conventional practices ($\approx D^{-0.0008}$).
- Accuracy Gains: At a matched budget of 40k samples, WiCompass reduced Mean Per Joint Position Error (MPJPE) by 25–30 mm.
- Real-World Feasibility: Live tests with bodyweight exercises confirmed that WiCompass-guided collection significantly outperforms standard action-list strategies.
Conclusion
WiCompass shifts the focus from brute-force data scaling to coverage-aware acquisition. By treating data collection as a search for missing "tokens" in a universal vocabulary, it provides a scalable path toward robust, privacy-preserving human sensing. Future work will aim to bridge the "sensing gap" caused by environmental factors like multipath interference.
Summary updated