acceptodds
Under review as a conference paper at ICLR 2027

From Hidden to Handy: Making Bias Explicit and Actionable in mmWave Pose Estimation

Abstract

Practical mmWave sensing systems operate with sparse returns, specular reflections, and occlusions, pushing pose estimators to lean on statistical priors rather than sensor evidence. This systems-level reality yields biased joint predictions that hurt downstream tasks (e.g., gesture/activity recognition). We present mmJoints, which augments a pre-trained, black-box mmWave 3D pose estimator with per-joint descriptors that characterize the degree of local radar support and the reliability of each predicted joint location. Instead of hiding bias, mmJoints makes it explicit and usable—improving interpretability of pose and accuracy of downstream models. Evaluated on over 115,000 frames across 13 pose-estimation settings, mmJoints consistently estimates informative and descriptors across diverse models and sensing conditions, improves joint accuracy by up to 12.5%, and boosts activity recognition accuracy by up to 16%.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
mmWave sensing, pose estimation, joint descriptor

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_coB8OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper addresses the issue of over-reliance on learned priors in mmWave-based human pose estimation due to sparse and unreliable radar observations. The authors propose MMJoints, a framework that augments pre-trained black-box pose estimators with two per-joint descriptors: Sensing Score and Reliability Score, which characterize local radar support and prediction reliability, respectively. To estimate these descriptors without ground-truth poses at inference time, the method leverages latent signal distribution modeling and surrogate pose estimation. The main contribution is a bias-aware pose representation that makes sensing support and prediction reliability explicit and actionable, improving downstream pose estimation and activity recognition without modifying the original pose estimator.

Critical strengths
  1. The paper addresses an important practical challenge in mmWave-based pose estimation, where sparse and unreliable radar observations can cause models to over-rely on learned human pose priors. Explicitly characterizing sensing support and prediction reliability at the joint level is meaningful, particularly for downstream applications where inaccurate or insufficiently supported predictions may lead to unreliable decisions.

  2. Generalizable Black-Box Framework. The proposed MMJoints framework operates as a post-hoc augmentation module without requiring access to or modification of the underlying pose estimator. This model-agnostic design makes it potentially applicable to different pre-trained architectures and facilitates integration into existing systems.

Critical weaknesses
  1. mmWave reflections originate from body surfaces rather than joint centers, making local radar returns an imperfect measure of sensing support. A joint predicted correctly despite sparse nearby returns may reflect effective information fusion from neighboring body parts, temporal context, and kinematic constraints, rather than harmful reliance on priors. Conversely, incorrect predictions despite strong nearby returns may arise from incorrect reflection-to-body-part associations. Although the proposed scores characterize local sensing support and prediction reliability, the paper does not clearly distinguish beneficial prior-based inference from harmful prior bias, leaving the interpretation of these mismatches insufficiently justified.
  2. The proposed method relies on learning a mapping from human poses to latent mmWave signal distributions. However, accurately modelling mmWave observations from human poses remains challenging due to complex reflections, multipath effects, and environmental variations, as highlighted by prior studies on mmWave signal synthesis (e.g., SynMotion, SenSys 2022). Although the authors model latent distributions rather than raw signals, it remains unclear whether the generated distributions can faithfully capture real-world radar observations. Consequently, the divergence between generated and observed distributions may reflect environmental variations rather than pose estimation errors, potentially undermining the reliability of the surrogate pose construction.
  3. The experimental comparisons are relatively limited, relying primarily on uncertainty estimation methods published between 2016 and 2021. Although the paper demonstrates improvements in downstream pose refinement and activity recognition, it lacks comparisons with recent downstream method. Especially Recent works, such as RadProPoser (IJCNN 2026) and UGSTM (2025), have already incorporated uncertainty modelling into radar-based pose estimation and skeleton-based action recognition, respectively. The paper would benefit from discussing these methods and, where experimentally feasible, comparing their uncertainty representations on downstream tasks to better establish the unique contribution of MMJoints.
  4. Since MMJoints is proposed as a model-agnostic add-on module, its applicability to more recent architectures remains insufficiently explored. Evaluating MMJoints with stronger skeleton-based action recognition models, such as SkateFormer (ECCV 2024) and FreqMixFormer (ACM MM 2024), or newer mmWave pose estimators such as Wave2Body (2026), would better substantiate its generalizability and practical benefits.
Questions for authors
  1. Can the authors provide controlled analyses showing how the descriptors distinguish beneficial use of contextual and kinematic priors from harmful prior-induced prediction errors?
  2. Can the authors validate the fidelity of the generated signal distributions under unseen sensing conditions? Additionally, does the distribution divergence reliably correlate with pose estimation errors rather than environmental variations?
  3. Can the authors provide controlled experiments demonstrating that the proposed descriptors distinguish harmful prior reliance from accurate predictions supported by contextual or kinematic information?
  4. Can the authors compare downstream performance using the original predicted poses augmented with MCD/DER/LULA uncertainty estimates versus MMJoints descriptors? Such comparisons would clarify whether the proposed descriptors provide additional benefits beyond existing uncertainty estimates.
  5. Can the authors compare MMJoints with recent uncertainty-aware methods and evaluate its integration with newer pose estimation and action recognition models to better demonstrate its effectiveness and generalizability?
  6. Can the authors clarify and compare the model sizes, training data, and supervision used by MCD, DER, LULA, and MMJoints to ensure a fair evaluation?
  7. Can MMJoints be extended to multi-person scenarios?
Reviewer_MhzSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles the fundamental limitations of 3D pose estimation problem on mmWave radar point clouds, which suffers from noisy and sparse observations due to limited antenna aperture. To address this, the authors propose MMJoints, a separately trained module that additionally estimates whether each estimated joint information is supported by actual radar evidence and how reliable it is. Plugging the proposed framework into baseline models across multiple benchmarks, the authors observe consistent improvements in downstream pose estimation and activity classification tasks.

Critical strengths
  • Noise and sparse observations in mmWave data are fundamental problems to solve in mmWave pose estimation, where the paper proposes an effective framework to solve this.
  • The proposed framework can easily be combined with existing mmWave pose estimation algorithms, which makes it highly practical and scalable.
  • Experiments across various benchmarks/baslines demonstrate the efficacy of the proposed approach, with consistent improvements on two downstream tasks.
Critical weaknesses
  • Even if the work is presented as a methodology for mmWave-based pose estimation, its core scheme appears largely independent of physical mmWave sensing. Namely, its main objective is to estimate the confidence of pose joints using surrogate model, without incorporating specific physical priors from mmWave radar data.
  • That is, the paper does not sufficiently explain why the proposed method is specific to mmWave radar.
  • If the framework is universal rather than specific to mmWave radar signals, the authors should demonstrate whether plugging the method into vision- or LiDAR-based pose estimators also lead to performance improvements.
  • Also, the framework is evaluated by combining it with the simple baseline models associated with the datasets. However, the models deployed are relatively basic and too outdated models. It is unclear whether the reported improvements remain valid even when the framework is applied to recent/SOTA mmWave point cloud models.
Questions for authors
  • Which parts of the proposed framework are explicitly designed to reflect the physical priors of the mmWave radar signals?
  • Can the framework be integrated with more recent mmWave-based pose estimators for further improving the pose estimation performance?
  • Does the model still remain effective even when applied to larger models? That is, is the framework scalable to larger models?
Reviewer_pxNwOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed to add two scores: sensing score and reliability score, to existing radar pose estimation pipelines. The sensing score measures the local radar support for predicted joints, while the realiability score quantifies the confidence in a joint estimate. The proposed pipeline learns a pose-conditioned Gaussian mixture model over radar features, and a surrogate-pose network. The surrogate network receives distribution-discrepancy features and the original pose estimate. A final network predicts the two descriptors.

Critical strengths

The perspective of identifying measurement supports for radar human pose estimation is interesting. It is quantified by using two proposed quantities such as the sensing and reliability scores. The designed pipeline appears to support and align this perspective.

Critical weaknesses
  1. The writing, experiment, and the result analysis appear to be very agentic with less human insights. The paper appears to be auto-generated by the agent. However, in the AI USE STATEMENT, I didn't find such a statement rather than AI assist for proofreading, language polishing, rephrasing, and improving the clarity and readability of the manuscript. They were also used for limited assistance with code editing and debugging. Also it is clear that the text didn't explain or match with Figures. For instance, Fig. 2, there is no annotation about signal encoder, rather than mmWave Frames...

  2. The two core metrics, sensing and reliability scores, lacks of deep understanding of radar signal processing.

2.a) Radar point clouds are processed in a way that they are subjected to many variables along the way, including physical resolution of the radar sensor, sampling rate, FFT or other high-resolution processing techniques, CFAR parameters, etc. These two scores appear to completely ignore these effects. 2.b) How will these two scores work for in the presence of multiple subjects? In the presence of false detection points (such as background reflection or clutter), in the presence of a ghost target? 2.c) Note that mm-Fi was collected mostly in the presence of a single subject in the scene and the subject is not moving in terms of depth, azimuth or elevation. The authors need to clarify how these two scores respond to point clouds when the number of point clouds (even for the same subject) changes significantly from near distance to far distance (in the extreme case, one point cloud per subject). 2.d) All the above points lead to the question how the pose/signal encoder learns in the case that the clusters of point clouds may depend on where the subjects are located in the 3D space. Similarly, how much is meaningful to consider the distribution in the latent space? Will different learned clusters correspond to the same pose but in different depths? 2.e) The GMM modeling for both pose and signal latent space needs to be clarified and better justified. This assumption significantly simplified the consequent analysis on divergence analysis.

  1. The evaluation over 3 different radar datasets: MARS, mmBody, and self-collected dataset. All three datasets appear to be single-subject at a time/frame and fairly open space. There exists many other radar datasets including heatmaps or raw data with single-subject and multiple subjects and in open/cluttered space, with free-form movements.
Questions for authors
  1. Sensing score: how is it defined in the case of multiple subjects, clutter, and multipath?
  2. How does the training-set reference mean for each joint j account for different depth and different subjects? Note that radar cross-range cell becomes larger when the depth is further. In other words, even for the same subject, the number of reflection points per subject is less when the subject is far from the radar. The training-set reference mean becomes averaged or smoothed if the depth-dependent effect is not properly accounted.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.