acceptodds
Withdrawn from ICLR 2027

ProxiHG: Latent Proximal-joint Hypergraph Encoding for Sparse-IMU 3D Human Pose Estimation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse-IMU-based 3D human pose estimation aims to recover full-body motion from a small number of inertial sensors. Under the standard six-IMU setting, key proximal joints, such as the shoulders and hips, are not directly observed. As a result, similar distal IMU signals may correspond to substantially different full-body poses, leading to inherent pose ambiguity. To address this issue, we propose ProxiHG, an encoder-side structural representation framework that explicitly introduces latent proximal joints before pose decoding. Specifically, ProxiHG employs a two-level hypergraph to first model relations among observable IMU nodes and then capture high-order dependencies among sensors, latent proximal joints, and body parts. The resulting structured representation is temporally fused and progressively decoded into full-body motion. Experiments on DIP-IMU and TotalCapture demonstrate consistent improvements in pose accuracy, spatial reconstruction, motion stability, and cross-dataset generalization. These results show that explicitly modeling unobserved proximal joints provides an effective structural inductive bias for sparse-IMU-based full-body motion estimation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sparse-IMU human pose estimation, 3D human motion reconstruction, hypergraph neural networks, latent structural modeling, wearable sensing, structured prediction

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_tX46OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates encoder-side structural representations for full-body pose estimation from sparse IMU measurements. It introduces latent shoulder and hip representations before pose decoding and models their higher-order interactions with observable sensors. The proposed method demonstrates promising performance in terms of reconstruction accuracy, temporal consistency, and cross-dataset generalization.

Critical strengths
  1. This paper introduces latent representations of the shoulders and hips before pose decoding, modeling their higher-order interactions with observed sensor inputs. 2)The proposed method demonstrates good performance in terms of reconstruction accuracy, temporal consistency, and cross-dataset generalization.
Critical weaknesses
  1. This paper does not compare with SOTA method in six-IMU configuration.
  2. cross-dataset generalization part is not clearly explained.
Questions for authors
  1. Why do the authors not compare the proposed method with state-of-the-art approaches under the same six-IMU configuration?

  2. Why is cross-dataset generalization evaluated only on the TotalCapture dataset, given its relatively small size? Evaluation on additional datasets would provide more convincing evidence of generalization.

  3. The ablation study should further investigate the contributions of different loss functions to better understand their impact on the overall performance.

Reviewer_erAdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on sparse-IMU-based 3D human pose estimation. Authors mainly addresses the pose ambiguity in this area. Specifically, they propose ProxiHG, an encoder-side structural representation framework that explicitly in-troduces latent proximal joints before pose decoding. Authors introduce hypergraph to capture human joints as well as high-order dependencies. Experiments are performed on DIP-IMU and To-talCapture.

Critical strengths

Authors present albation studies.

Critical weaknesses
  • What are the specific advantages of using a hypergraph? Does the hypergraph propagation proposed by the authors appear to be a weighted aggregation of ordinary graphs? The experimental results also indicate that the hypergraph provides only a very limited performance improvement, such as 0.2 Ang.
  • The authors use GRU for information propagation. Why not use the Transformer architecture directly, which is very strong at handling time series data?
  • Please provide visualizations of the potential shoulder and hip node features to explicitly assess their structural semantics.
  • Lack failure case analysis.
Questions for authors
  • Please clarify the motivation and effects of hypergraph more clearly.
  • Please provide more experiments on different architectures.
  • Please provide intermediate feature visualizations.
Reviewer_ai3nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ProxiHG estimates full-body pose from six IMUs. It models the observed sensors and adds four latent shoulder and hip nodes, connected through body-part hyperedges. Temporal fusion and a staged decoder then predict joint positions, rotations, and translation. The experiments use DIP-IMU and test transfer from AMASS to real TotalCapture measurements, with the TotalCapture subset removed from AMASS training.

Critical strengths
  1. Adding shoulder and hip features between the distal sensors and torso makes sense for this sensor layout. Table 2 gives a useful control: adding the latent component to Star+Part reduces position error from 4.45 to 4.14 cm. SIP measures the four shoulder and hip joints, and Figure 5 shows a related pose example and frame-wise error curves.

  2. The TotalCapture test is a strength. The model sees no TotalCapture data during training or model selection and is tested directly on real IMU measurements. The appendix is also clear about preprocessing and the offline evaluation protocol.

Critical weaknesses
  1. I am not convinced that Table 3 establishes a distinct high-order modeling advantage. The propagation uses node-to-node matrices: the observed matrix is learned after hypergraph initialization, and the latent matrix is fixed. A graph using the same propagation operator and remaining operations would produce the same output. I would therefore like to understand whether the graph controls differ in weighting, normalization, or learnability. The four latent nodes have an anatomical assignment, but it is also unclear whether they help more than generic latent features of the same size.

  2. The appendix clearly states that the bidirectional decoder uses future frames within each five-second window. What is missing is the corresponding setting for each baseline. This affects how I interpret the accuracy and JVE comparisons; the within-model structural ablations keep the decoder fixed. Also, position, mesh, and velocity errors are root-aligned, so the reported results do not evaluate global translation. The claims about global motion recovery should be narrowed accordingly.

Questions for authors
  1. How are the graph controls in Table 3 constructed? Please give their propagation matrices, normalization, learnable weights, and parameter counts, and explain how these differ from the hypergraph versions.
  2. How much future information and what inference window does each baseline in Table 1 use? Please identify the retrained models and the PNP official checkpoint separately.
  3. Have you tested four generic latent nodes with comparable capacity, or a setting with the anatomical correspondence disrupted? Either would help explain the benefit of the shoulder/hip assignment beyond adding latent features.

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

Related papers

Discussion (0)

Sign in to comment.