cs.CVMar 3, 2026

Synthetic-Child: An AIGC-Based Synthetic Data Pipeline for Privacy-Preserving Child Posture Estimation

Authors: Taowen Zeng

Organizations: Independent Researcher

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Accurate child posture estimation is critical for AI-powered study companion devices, yet collecting large-scale annotated datasets of children is both expensive and ethically prohibitive due to privacy concerns. We present Synthetic-Child, an AIGC-based synthetic data pipeline that produces photorealistic child posture training images with ground-truth-projected keypoint annotations, requiring zero real child photographs. The pipeline comprises four stages: (1) a programmable 3D child body model (SMPL-X) in Blender generates diverse desk-study poses with IK-constrained anatomical plausibility and automatic COCO-format ground-truth export; (2) a custom PoseInjectorNode feeds 3D-derived skeletons into a dual ControlNet (pose + depth) conditioned on FLUX-1 Dev, synthesizing 12,000 photorealistic images across 10 posture categories with low annotation drift; (3) ViTPose-based confidence filtering and targeted augmentation remove generation failures and improve robustness; (4) RTMPose-M (13.6M params) is fine-tuned on the synthetic data and paired with geometric feature engineering and a lightweight MLP for posture classification, then quantized to INT8 for real-time edge deployment. On a real-child test set (n~300), the FP16 model achieves 71.2 AP -- a +12.5 AP improvement over the COCO-pretrained adult-data baseline at identical model capacity. After INT8 quantization the model retains 70.4 AP while running at 22 FPS on a 0.8-TOPS Rockchip RK3568 NPU. In a single-subject controlled comparison with a commercial posture corrector, our system achieves substantially higher recognition rates across most tested categories and responds ~1.8x faster on average. These results demonstrate that carefully designed AIGC pipelines can substantially reduce dependence on real child imagery while achieving deployment-ready accuracy, with potential applications to other privacy-sensitive domains.

Figures & tables

Figure 1: Overview of the Synthetic-Child pipeline. Stage 1 : A child-specific SMPL-X model in Blender generates diverse desk-study poses with IK constraints, exporting rendered images, depth maps, and COCO-format keypoint annotations. Stage 2 : A custom PoseInjectorNode feeds ground-truth skeletons into a dual ControlNet (pose + depth) conditioned on FLUX-1 Dev, producing 12,000 photorealistic child images with low annotation drift. Stage 3 : Three automated quality gates—confidence filtering (ViTPose-H), spatial fidelity checking, and category consistency verification—remove generation failures and yield 11,900 curated images from an initial pool of 12,000; online augmentation further increases robustness. Stage 4 : RTMPose-M is fine-tuned on the synthetic data; geometric feature engineering and a lightweight MLP classify postures; the model is quantized to INT8 for real-time inference on an edge NPU.
IDCategoryDescription
0correct_postureUpright spine, shoulders level, eyes forward
1bowed_headModerate forward head tilt ( ∼15{\sim}15 – 30∘30^{\circ} )
2very_bowedSevere forward flexion ( >30∘>30^{\circ} )
3lean_deskUpper body collapsed onto desk surface
4lean_leftLateral trunk lean to the left
5lean_rightLateral trunk lean to the right
Table 1: The C=10C{=}10 posture categories generated in Stage 1. Each template specifies a canonical joint configuration; stochastic perturbation (Eq. 2 ) generates diverse variants within each category.
Figure 2: Illustration of the Stage 1 →\to Stage 2 pipeline for two representative samples viewed from a frontal desk-mounted camera perspective. Row 1 : correct upright posture; Row 2 : upper body collapsed onto desk ( lean_desk ). Each row shows: (a) Blender render with overlaid 3D skeleton, (b) depth map rendered from the Blender Z-buffer, (c) COCO-17 OpenPose skeleton deterministically rendered by PoseInjectorNode from the ground-truth JSON, and (d) photorealistic image generated by FLUX-1 conditioned on both control signals (pose + depth). The stark visual contrast between (a) and (d) demonstrates the pipeline’s ability to bridge the sim-to-real gap while faithfully preserving the input pose.
Training DataModelParamsAP[email protected]FPS
  (M)  (NPU)
COCO pre-trainedRTMPose-M13.658.771.418
Synthetic-ChildRTMPose-M13.671.284.518
Synthetic-Child (INT8 †\dagger )RTMPose-M13.670.483.722
Table 2: Keypoint estimation accuracy on the real-child test set ( n≈300n{\approx}300 ). All baseline and proposed models use the RTMPose-M architecture (13.6M params). “ †\dagger ” denotes INT8 quantized model; all other rows use FP16 precision. All models are deployed on the same RK3568 NPU. Best results are in bold .
Posture BehaviorAngle/ConditionnnOursCommercialΔ\boldsymbol{\Delta}
Head turning15∘15^{\circ} – 50∘50^{\circ} ( 5∘5^{\circ} steps)2487.5%83.3%+4.2%
Head too low20∘20^{\circ} – 60∘60^{\circ} ( 10∘10^{\circ} steps)1586.7%6.7%+80.0%
Lateral leaning10∘10^{\circ} – 40∘40^{\circ} ( 5∘5^{\circ} steps)2166.7%47.6%+19.1%
Body rotation15∘15^{\circ} – 50∘50^{\circ} ( 5∘5^{\circ} steps)2491.7%——
Lying on deskFully collapsed3100%100%0%
Table 3: Pilot recognition rate (RR) comparison with a commercial RGB-based posture corrector (single subject, single environment). RR is the proportion of test instances where the device triggers a correct alert within 30 s. “—” indicates the device does not support this detection category.
Posture BehaviornnOurs (s)Commercial (s)Speedup
Head turning244.7110.202.16×2.16\times
Head too low159.0810.001.10×1.10\times
Lateral leaning216.5013.702.11×2.11\times
Body rotation245.80——
Lying on desk36.3310.001.58×1.58\times
Weighted avg. (shared)636.4211.311.76×1.76\times
Table 4: Average response time TavgT_{\text{avg}} (seconds) from posture onset to alert trigger. Per-category averages are computed over successfully detected instances; nn denotes the total number of test instances per category (same as Table 3 ). Lower is better. “—” indicates the device does not support this category. The weighted average is computed over shared categories only.
ConfigurationAP[email protected]Δ\boldsymbol{\Delta} AP
Full Synthetic-Child pipeline71.284.5—
(a) Synthesis configuration   
w/o depth ControlNet ( w2=0w_{2}{=}0 )68.181.5−- 3.1
(b) Quality control   
w/o ViTPose confidence filtering67.881.2−- 3.4
w/o data augmentation69.182.6−- 2.1
Table 5: Ablation study on key pipeline components. All variants use RTMPose-M evaluated on the real-child test set. Δ\Delta AP denotes the change relative to the full pipeline.
Figure 3: Qualitative comparison on challenging child postures from the real-child test set, selected for their difficulty under the COCO-pretrained baseline. Each pair shows the baseline output ( left : estimated keypoint skeleton only) and the full Synthetic-Child pipeline output ( right : keypoints + detection bounding box + MLP classification label with confidence score). (a, b) Two subjects leaning onto the desk surface with ≥{\geq} 0.97 confidence, a common desk-study posture where children’s compressed upper body and foreshortened limbs diverge markedly from adult training distributions. (c) Rightward lateral trunk lean (0.96 confidence). (d) Correct upright posture (0.91 confidence), included to show that the pipeline correctly identifies proper sitting as non-deviant rather than triggering false alerts. The baseline provides keypoint estimation only and cannot classify posture; our pipeline enables category-level posture alerts with high confidence.
More in this topic

Human Pose Estimation

Human pose estimation recovers people’s body keypoints, 3D skeletons, or meshes from images, video, and other sensor data.

See all 74 papers

Explore similar work

CardsList
Aug 6, 2026

Patient Pose Assessment Using a CT-Based Framework for Synthetic Data Generation

An adequate diagnostic quality of radiographs is essential for reliable diagnoses and treatment planning. The patient's pose during radiography is one of the most important factors determining the diagnostic quality. Since patient positioning is difficult and not standardized, an automated AI-based approach using depth images to automatically assess the patient's pose before the radiograph has been taken would be helpful. Due to regulatory hurdles, however, it is difficult in practice to acquire the required depth images and corresponding radiographs. In this paper, we present a framework that can generate such training data synthetically from Computed Tomography scans. We further show that by pretraining on our generated synthetic dataset consisting of 3077 image pairs of upper ankle joints, the pose assessment of real upper ankle joints can be improved by up to 11 percentage points.

Manuel Laufer, Dominik Mairhöfer, Malte Sieren +7

Institute for Neuro- and Bioinformatics, University of Lübeck, Germany · University Medical Center Schleswig-Holstein, Lübeck, Germany · IMAGE Information Systems Europe GmbH, Rostock, Germany +1

Synthetic Data PretrainingMedical ImagingSynthetic Data GenerationMedical Image Synthesis
AskPDF
Mar 30, 2026

PoseDreamer: Scalable and Photorealistic Human Data Generation Pipeline with Diffusion Models

Acquiring labeled datasets for 3D human mesh estimation is challenging due to depth ambiguities and the inherent difficulty of annotating 3D geometry from monocular images. Existing datasets are either real, with manually annotated 3D geometry and limited scale, or synthetic, rendered from 3D engines that provide precise labels but suffer from limited photorealism, low diversity, and high production costs. In this work, we explore a third path: generated data. We introduce PoseDreamer, a novel pipeline that leverages diffusion models to generate large-scale synthetic datasets with 3D mesh annotations. Our approach combines controllable image generation with Direct Preference Optimization for control alignment, curriculum-based hard sample mining, and multi-stage quality filtering. Together, these components naturally maintain correspondence between 3D labels and generated images, while prioritizing challenging samples to maximize dataset utility. Using PoseDreamer, we generate more than 500,000 high-quality synthetic samples, achieving a 76% improvement in image-quality metrics compared to rendering-based datasets. Models trained on PoseDreamer achieve performance comparable to or superior to those trained on real-world and traditional synthetic datasets. In addition, combining PoseDreamer with synthetic datasets results in better performance than combining real-world and synthetic datasets, demonstrating the complementary nature of our dataset. We will release the full dataset and generation code.

Lorenza Prospero, Orest Kupyn, Ostap Viniavskyi +2

The Podium Institute for Sports Medicine and Technology University of Oxford Oxford, UK · Visual Geometry Group University of Oxford Oxford, UK · Ukrainian Catholic University Lviv, Ukraine

Synthetic Data GenerationHuman Mesh RecoveryDiffusion-Based Image Generation
AskPDF
Jul 14, 2026

Metric-Guided Synthetic Image Data Rendering for Deep Learning compatible with Agentic AI

Deep learning computer vision for scientific applications requires collecting and annotating large datasets in a laborious, expensive and error-prone process. Synthetic data generation through 3D modelling and rendering may simplify this process and increase the accuracy of annotations by generating them programmatically. However, minimising the domain gap between real and synthetic images visually is subjective and lacks systematic quantitative guidance. We present GraNatPy, a Python package with metrics to guide improvement of the rendered scene. We show that quantifiable increase in realism, diversity and size of rendered dataset correlates with improved visual perception of the scene and higher zero-shot performance of an object detection model. Furthermore, we demonstrated using photographs of virological plaque assays that gradient similarity affects performance on small object detection, which can be improved by mixing real and synthetic data. Finally, we turn procedural data rendering into an agentic skill (SynthClaw) to automate the procedural parameter optimisation.

Martina Radoynova, Samuel Pantze, Trina De +2

Center for Advanced Systems Understanding (CASUS), Görlitz, Germany · Helmholtz-Zentrum Dresden-Rossendorf e. V. (HZDR), Dresden, Germany · Institute of Computer Science, University of Wrocław, Wrocław, Poland +1

Synthetic-to-Real Domain AdaptationSynthetic Data GenerationSynthetic Data Evaluation
AskPDF