acceptodds
Withdrawn from ICLR 2027

LDBody: Benchmarking Human Pose and Mesh Recovery with Residual Limb and Prostheses

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Human pose and mesh benchmarks largely assume limb-complete bodies. This leaves residual anatomy, prosthesis-supported motion, and variation in body pro- portions insufficiently evaluated. We introduce LDBODY, a dataset for pose and 3D human mesh evaluation. To our knowledge, it is the first dataset to provide cali- brated human body mesh evaluation with separate references for residual anatomy and prosthesis-supported function. Its annotation framework distinguishes body joints, residual-limb endpoints, and prosthesis-supported landmarks. Multiple internationally accredited Para Athletics classifiers guide the annotation protocol. The dataset contains an in-the-wild image component, LDBODY Images, and a calibrated motion component, LDBODY Motion, with shared landmark definitions. The image component covers upper- and lower-limb differences and varied body proportions, activities, scenes, clothing, and viewpoints. It adds prosthesis annota- tions to 8,866 LDPose images while retaining existing body-joint and residual-limb labels. The motion component captures real movements with eight synchronized RGB cameras and Vicon. It covers multiple prosthesis types and conditions with and without prostheses. The recordings provide 1,088 view-specific videos and 963,104 RGB frames. Vicon-derived keypoints and Vicon-constrained fitted body meshes support separate evaluation of residual anatomy and prosthesis-supported function. We benchmark 2D pose, multi-view 3D pose, residual-limb localization, and human mesh recovery. Additional analyses examine movement, viewpoint, and prosthesis conditions. Results on the evaluated limb-deficiency cases reveal persistent errors at residual endpoints and affected distal joints. Similar joint errors can also coexist with large differences in fitted-surface agreement. LDBODY sup- ports research on accessible motion analysis, assistive technologies, and inclusive sport by making diverse bodies explicit evaluation targets.

Primary area:
datasets and benchmarks
Keywords:
Individuals with Limb Deficiency, Dataset

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

This paper introduces LDBODY, a benchmark for 2D/3D human pose estimation and mesh recovery for individuals with limb deficiencies and prosthesis users. The dataset combines in-the-wild images with newly collected multi-view motion data captured using eight synchronized RGB cameras and a Vicon system. It provides annotations for body joints, residual-limb endpoints, and prosthesis-supported landmarks, along with fitted mesh references. The authors also evaluate several existing pose estimation and mesh recovery methods and analyze their performance across different movement and prosthesis conditions.

Critical strengths

The authors created a dataset for 2D and 3D pose estimation in prosthesis-related scenarios and also provided fitted mesh reference data. They provide both an in-the-wild dataset and a calibrated dataset, which complement each other to some extent. Specifically, the in-the-wild images are based on the existing LDPose dataset with additional annotations, while the calibrated motion data were newly collected by the authors. The evaluation metrics are generally standard and widely recognized, relatively comprehensive, and reasonably adapted to prosthesis-related scenarios.

Critical weaknesses

My main concern is the limited number of subjects in LDBODY Motion. The dataset only includes five participants, which is insufficient for a benchmark, especially one intended to evaluate generalization across different body structures and prosthesis configurations. Furthermore, all five participants are male, and four are active Para athletes. This raises concerns about the lack of diversity and how representative the benchmark really is. I consider this a major limitation of the dataset.

More importantly, with such a small number of subjects, it is difficult to know whether the reported results reflect the general performance of the evaluated methods or are strongly influenced by the characteristics of these particular individuals. The large number of recorded frames does not really address this issue, since they all come from the same five participants. While the in-the-wild image dataset provides additional diversity, it is mainly an extension of the existing LDPose dataset and does not compensate for the limited diversity of the calibrated 3D data.

I think the dataset is a useful contribution, but its current scale makes it difficult to establish a sufficiently representative and reliable benchmark for broader evaluation.

Questions for authors

How do the authors see the benchmark supporting generalization across different populations and prosthesis configurations since there are only 5 subjects? Are there any plans to expand the dataset to include more diverse participants, particularly female and non-athlete prosthesis users?

Reviewer_8Q3EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LDBody is a dataset and benchmark for 2D pose, multi-view 3D pose, residual-limb endpoint localization and human mesh recovery for people with limb deficiency. It has two parts: LDBody Images adds prosthesis landmarks to 8,866 LDPose in-the-wild images. LDBody Motion has 136 tests from five adult male participants with lower-limb deficiency. For mesh evaluation, the authors fit two pseudo-ground-truth references per frame from HMR initializations refined to Vicon keypoints. Results on the evaluated limb-deficiency cases reveal persistent errors at residual endpoints and affected distal joint.

Critical strengths
  1. It fills a real evaluation gap. Standard pose and mesh benchmarks implicitly assume limb-complete bodies. LDBody provides the first calibrate evaluation that separates residual anatomy from prosthesis-supported function.

  2. The authors make responsible and transparent data practice.

Critical weaknesses
  1. LDBody part only covers five adult men, all with lower-limb deficiency. Therefore, it might not be diverse enough to benchmark 3D human pose and mesh recovery with residual limbs and prostheses.

  2. The mesh references and the evaluation protocol may not be accurate enough for a benchmark. The ground truth values are extracted by refining HMR outputs with Vicon joints. They keep 31–48 mm mean joint residuals and 16–85 mm endpoint error depending on the initializer (Table 12).

  3. The primary contribution of this paper appears to lie in dataset construction and benchmarking rather than algorithmic innovation.

Questions for authors

same as weaknesses

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

This paper introduces LDBody, a benchmark dataset for human pose estimation and mesh recovery for individuals with limb deficiencies and prosthesis users. Building on LDPose, the authors add prosthesis-related annotations and collect motion data for evaluation. Using these data, they construct 2D and 3D keypoint references and benchmark existing methods for 2D pose estimation, multi-view 3D pose estimation, residual-limb localization, and human mesh recovery. They further analyze performance across participants, movements, and prosthesis conditions.

Critical strengths
  1. The work addresses a gap in existing annotation schemes, which do not adequately cover individuals with limb deficiencies. It provides a useful foundation for future research on visual understanding involving this population.

  2. The authors conduct extensive evaluations of existing methods in this setting. These evaluations help identify the limitations of current approaches to pose estimation and mesh recovery for individuals with limb deficiencies.

Critical weaknesses

The contribution beyond prior work and the representativeness of the newly collected data are limited. The image component mainly adds prosthesis-related annotations to LDPose. Although the newly collected motion data enable 3D evaluation, the cohort consists of only five adult male participants, with some prosthesis types represented by a single participant. Despite the participant-level analyses, this small cohort limits coverage of individual differences, body configurations, and prosthesis types, leaving it unclear how well the evaluation findings generalize to a broader population.

The reliability of the residual-endpoint references when markers are missing requires further validation or clarification. The authors could provide a more detailed explanation of how missing markers affect the construction and reliability of these references.

Questions for authors

Could the authors provide more information on the variation in body shape and other relevant characteristics among the five participants? To what extent does this cohort represent the broader range of body configurations, limb deficiencies, and prosthesis types targeted by the benchmark?

What is distinctive about the additional annotations compared with those already provided by LDPose, and what new evaluation capabilities or insights do they enable? Please clarify how these annotations, together with the newly collected 3D data, constitute a substantive standalone contribution rather than an incremental update to LDPose

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

Related papers

Discussion (0)

Sign in to comment.