acceptodds
Withdrawn from ICLR 2027

SynplerHuman: Simplified Synthetic Data Generation for Estimating 3D Human Pose and Shape

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Synthetic training data is widely used for 3D human pose and shape estimation and recent datasets invest large efforts in increasingly realistic humans, clothing, and environments. However, these improvements are usually introduced together, making it difficult to isolate which factors matter most for real-world generalization. To remedy this problem, we present SynplerHuman, a controllable, open-source procedural generator of data for 3D human pose and shape estimation". Our generator procedurally generates bodies, clothing, materials, and scene variations, enabling unlimited data generation without relying on external curated 3D appearance asset libraries. Using SynplerHuman, we study what factors of realism matter in synthetic data generation. We find that articulation choices have the largest impact on performance, while appearance variations produce relatively small changes. Finally, we demonstrate that SynplerHuman produces data that is effective when combined with existing datasets, and is competitive with synthetic datasets such as SynBody.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
human pose estimation, synthetic human data

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

SynplerHuman is a controllable procedural pipeline for generating synthetic supervision for 3D human pose and shape estimation. Built on Infinigen, Blender, and SMPL-X, it varies appearance, articulation, grounding, and global orientation within a common generation framework. The paper introduces SynplerHuman-600k, described in Section 3.2 as 600,000 annotated human instances across 4,780 scenes. Using a CameraHMR model adapted to SMPL-X, the authors evaluate transfer to EMDB, 3DPW, SPEC, and RICH. The main finding is that replacing the learned pose prior substantially degrades performance, whereas several appearance modifications have smaller effects. Dataset comparisons and mixing experiments further demonstrate utility relative to SynBody and as a complement to BEDLAM2.

Critical strengths

1.A useful framework for investigating synthetic-data design. The pipeline exposes appearance, articulation, and physical-placement controls within one system, making comparisons more interpretable than comparisons between independently constructed datasets. Procedural clothing and scene generation could also reduce dependence on manually curated assets.

2.Strong evidence for the importance of the chosen pose distribution. Across four benchmarks, learned-prior sampling substantially outperforms the tested independent and rule-based alternatives. Increasing rule-based data scale and model capacity narrows but does not close the gap. These large differences provide actionable guidance within the studied training setup, and the dataset comparisons demonstrate that the generated supervision is practically useful.

Critical weaknesses

The broader design recommendation is insufficiently validated across learning setups. All core experiments use CameraHMR with the same pretrained ViTPose initialization. Existing appearance invariance may therefore contribute to the small appearance effects. Moreover, the pose interventions change distributional support and inter-joint correlations, not only physical realism. The evidence establishes the superiority of the selected learned prior over the tested samplers, but does not yet establish how broadly articulation should be prioritized over appearance.

Small-effect conclusions lack statistical support. In Table 4, floating humans improve PA by just 0.09 mm while slightly worsening MPJPE and PVE, making the claimed advantage unconvincing without uncertainty estimates. The core per-variant sample counts, training duration, and benchmark-specific SMPL-X joint/mesh evaluation mappings are also insufficiently specified, limiting verification of the controlled comparisons.

Questions for authors
  1. Does the relative sensitivity to appearance versus articulation persist under a different backbone or pretraining initialization? A focused comparison of representative variants would help distinguish a general data-design finding from an effect specific to pretrained CameraHMR.

  2. Section 3 states that the baseline reuses BEDLAM skin textures. Which released configurations are entirely independent of curated appearance assets, and what dependencies remain? Please also clarify whether “600k” counts rendered images or human instances/crops, and provide the per-variant breakdown.

  3. What are the exact sample counts and training steps for each core variant? Are scene, camera, body-shape, and sampling seeds paired across variants? Please specify the collision-filter threshold and the joint/mesh mappings and alignment procedures used on each benchmark.

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

The paper introduces a procedural pipeline for synthetic human data which studies how different factors (skin textures, clothing, and pose prior, etc) affect HMR accuracy. The proposed SynplerHuman dataset is compared with existing synthetic datasets such as Synbody and Bedlam. The research question is interesting and SynplerHuman outperforms some representative synthetic datasets (SynBody and AGORA), but still lags behind BEDLAM2.0. However, the cross-dataset comparisons use relatively small training subsets and are evaluated only one training model and on one test dataset(EMDB).

Critical strengths
  1. The paper addresses an interesting question: how different components of a synthetic human datasets (appearance, articulation, and physical plausibility) affect HMR performance. The use of a controllable procedural generator provides a promising framework for studying these factors.

  2. Under a controlled experimental setting, where CameraHMR is trained on different synthetic data and evaluated on EMDB, SynplerHuman outperforms representative synthetic datasets (SynBody and AGORA).

Critical weaknesses
  1. The experimental comparison with existing synthetic datasets is insufficient.

The comparison with existing synthetic datasets is conducted with only one model, CameraHMR, and on only one benchmark, EMDB. The authors sample approximately 101K crops from each dataset. However, SynBody contains 1.2M images and BEDLAM2.0 contains 8M images. The number of human crops available for training is even larger, so the comparison uses only a small part of the data of the related works. The sampled BEDLAM2.0 subset performs worse than the result reported for BEDLAM2.0 under CameraHMR training and EMDB evaluation setting. This gap highlights the need to clarify how the reduced training set affect performance. It is also unclear how sensitive the reported ranking is to the choice and number of the training samples. It is therefore difficult to assess whether the reported performance ranking would remain consistent with different subsets, model architectures, test benchmarks, or training scales.

  1. The analysis does not sufficiently explain the gains over existing synthetic datasets.

Table 3 shows that SynplerHuman outperforms SynBody on EMDB. Table 4 further shows that, starting from 35.9K BEDLAM2.0 crops, adding 11K crops from the SynplerHuman baseline or some of its variants yields better performance than adding the same number of BEDLAM2.0 crops. These are interesting findings, but the discussion largely restates the performance differences without sufficiently investigating what produces them. SynBody and BEDLAM2.0 offer richer appearance assets and rendering detail, including more varied hairstyles, clothing styles, and footwear, and also draw on large-scale real motion datasets such as AMASS. Given these strengths, the advantages demonstrated by SynplerHuman warrant a clearer explanation. The paper does not adequately identify which properties of its data enable models to achieve better results in these settings.

Questions for authors

The paper presents an interesting study of how different factors affect single-frame HMR accuracy. The finding that pose sampling has a larger impact than appearance is interesting, but the analysis offers limited further insight into how to design a better synthetic training dataset. Besides, comparisons with existing datasets use relatively small subsets. It therefore remains unclear whether the dataset ranking and the conclusions about the importance of different factors would hold at larger training scales. My score could increase if the authors address the following questions with supporting evidence:

  1. Can the authors explain which specific properties of SynplerHuman account for its advantage over SynBody and support this explanation with convincing analysis?

  2. Can the authors extend the comparisons with other synthetic datasets to larger training sets and additional test benchmarks?

Reviewer_SzJ8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SynplerHuman, a controllable procedural generation framework for studying how synthetic data design affects 3D human pose and shape estimation. The framework generates synthetic images of humans, clothing, and 3D scenes with corresponding annotations, and creates dataset variants by varying appearance, pose sampling, global orientation, and ground contact. Comparisons with existing synthetic datasets and mixed-data training experiments demonstrate the utility of the generated data for training.

Critical strengths
  1. The generation framework offers practical value by integrating humans, clothing, and scenes, reducing reliance on costly manually created assets.

  2. The experiments provide insights into data design that can help identify which factors to prioritize when generating synthetic data.

Critical weaknesses
  1. The paper mainly contributes a procedural data generation framework, a dataset, and a systematic evaluation. However, all downstream experiments use CameraHMR, so it is unclear whether the findings about different data design choices also hold for other human pose and shape estimation methods. Testing a few representative models with different architectures or training setups would make the conclusions more convincing.

  2. The diversity of the generated data could be examined more closely. Although the dataset contains many procedural scenes, the paper provides little information about scene categories or the types and severity of occlusion. It also lacks a breakdown of performance under challenging conditions, such as heavy occlusion. This makes it hard to tell how well the dataset covers real-world challenges and whether the findings still hold in these cases.

Questions for authors
  1. Could you run experiments with models other than CameraHMR, or give some intuition about how other frameworks might benefit from the generated data?

  2. Could you provide more details about the generated data, particularly the scenes and occlusions it covers, and show whether the main findings still hold in hard cases?

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

Related papers

Discussion (0)

Sign in to comment.