Source-linked AI summary
SIZER: A Dataset and Model for Parsing 3D Clothing and Learning Size Sensitive 3D Clothing
Garvita Tiwari, Bharat Lal Bhatnagar, Tony Tung, Gerard Pons-Moll
TL;DR
The paper addresses the absence of real-data models that predict clothing deformation as garment size changes. It introduces the SIZER dataset, SizerNet for size-conditioned draping, and ParserNet for single-pass mesh parsing, reporting better parsing accuracy and size prediction than baselines trained on SIZER.
Problem
Existing real-data clothing models do not predict garment deformation as a function of garment size, and suitable real scans with size variation are lacking.
Method
The paper introduces SIZER, SizerNet for body- and size-conditioned garment draping, and ParserNet for converting a registered mesh into layered garments and body shape.
Results
Experiments show better clothing parsing accuracy and size prediction than baseline methods trained on SIZER.
Takeaways & Limitations
The combined models enable visualizing garments in different sizes and editing clothing directly on input meshes without scan segmentation.
Abstract
from arXiv · showhide
While models of 3D clothing learned from real data exist, no method can predict clothing deformation as a function of garment size. In this paper, we introduce SizerNet to predict 3D clothing conditioned on human body shape and garment size parameters, and ParserNet to infer garment meshes and shape under clothing with personal details in a single pass from an input mesh. SizerNet allows to estimate and visualize the dressing effect of a garment in various sizes, and ParserNet allows to edit clothing of an input mesh directly, removing the need for scan segmentation, which is a challenging problem in itself. To learn these models, we introduce the SIZER dataset of clothing size variation which includes $100$ different subjects wearing casual clothing items in various sizes, totaling to approximately 2000 scans. This dataset includes the scans, registrations to the SMPL model, scans segmented in clothing parts, garment category and size labels. Our experiments show better parsing accuracy and size prediction than baseline methods trained on SIZER. The code, model and dataset will be released for research purposes.
1 Introduction
The paper addresses the lack of real-scan data and models for predicting clothing fit across garment sizes and body shapes. It introduces SIZER, SizerNet, and ParserNet to learn size-sensitive draping and directly parse clothed meshes.
- Motivation: The problem matters because clothing fit is nonlinear and subtle, yet online purchases can be difficult without knowing how garments will fit.The paper connects this challenge to applications including virtual try-on, content generation, and clothing recommendation.
- Dataset: SIZER provides approximately 2000 scans of 100 subjects wearing 10 garments in four sizes, with registrations, segmentations, body shapes, and size labels.The dataset supplies the multi-layer annotations needed to learn clothing variation from real scans.
- SizerNet: SizerNet predicts how a garment drapes on a body as a function of body shape and input and desired garment sizes.Its decoder receives body-shape parameters and one-hot encodings of the input and output sizes.
- ParserNet: ParserNet maps one SMPL-registered mesh to layered garments and body shape under clothing in a single feed-forward pass, avoiding scan segmentation and nonlinear optimization.It also reparameterizes surfaces to match common garment templates while preserving personal details.
- Applications: The proposed system enables visualizing dressing effects across sizes and editing garment meshes directly on existing input meshes.The tandem of SizerNet and ParserNet supports separate garment and body-shape manipulation.
2 Related Work
Prior work models clothing from pose, body shape, images, simulations, or segmented scans, but existing datasets and methods do not provide realistic, annotated size variation across real clothed subjects.
- Clothing modeling: Existing clothing models learn deformation from pose or body shape, while simulation-based style spaces target plausibility rather than deformation of real garments on real bodies.The related methods therefore do not directly address size-conditioned real-garment draping.
- Shape under clothing: Body-shape estimation methods fit models under clothing or learn from images, but complex or loose garments remain difficult without additional priors or supervision.These approaches focus on recovering the body rather than providing size-varying garment layers.
- Cloth parsing: 3D cloth-parsing methods segment clothing from scans using multi-step procedures such as MRF-GrabCut and transfer the results across subjects.The paper positions ParserNet as a single-pass alternative to this procedural parsing workflow.
- 3D datasets: Existing 3D datasets are synthetic, smooth, minimally clothed, or lack garment segmentation, and none provides substantial real-scan clothing-size variation with corresponding body shapes.The cited real datasets also differ in scale, capture setting, or annotation coverage.
3 Dataset
SIZER is a real-scan dataset designed to support modeling garment fit across body shapes and sizes, with registrations and labels enabling standardized analysis and downstream clothing applications.
- 3 Dataset: Over 2000 scans cover 100 subjects wearing 10 garment classes in multiple sizes, addressing the lack of calibrated real-world size-variation data.Subjects wore the same garments in two or three sizes, including S, M, L, and XL.
- 3 Dataset: The dataset provides scans, segmented clothing, SMPL+G registrations, garment categories, and size labels for each scan.These annotations support correspondence, garment analysis, and size-sensitive modeling.
- 3 Dataset: SMPL models body pose and shape with deformation offsets, while SMPL+G represents the body and garments as separate registered meshes.Garment templates are defined as subsets of the SMPL template with garment-specific connectivity.
- 3 Dataset: The released dataset, code, and baseline models support virtual try-on, character animation, generative modeling, body-shape estimation, garment parsing, and garment resizing.The dataset is intended as a benchmark for 3D clothing parsing and garment resizing.
4 Method
The method combines ParserNet for single-mesh clothing and body-detail parsing with SizerNet for body- and size-conditioned garment resizing, trained using geometric and regularization losses.
- 4.2 ParserNet: ParserNet predicts garment layers and detailed body shape from a single registered mesh, avoiding separate segmentation and nonlinear optimization.It first estimates SMPL pose and shape, then extracts garment layers and personal features using multiple subnetworks.
- 4.1 Pose and shape prediction network: Body-shape prediction estimates undressed SMPL pose and shape from a clothed mesh using parameter and per-vertex losses, producing a smooth body under clothing.An additional surface loss brings the predicted undressed body closer to the clothed input mesh.
- 4.2 ParserNet: ParserNet predicts a sparse regressor that forms each garment vertex as a convex combination of nearby input-mesh vertices while preserving fixed garment-template connectivity.This converts instance-specific scan geometry into coherent garment meshes with shared class-specific topology.
- 4.3 SizerNet: SizerNet encodes an input garment into a latent code and conditions decoding on body shape plus one-hot input and desired garment sizes.The decoder predicts garment offsets on a template and produces the garment in the requested output size.
- 4.4 Loss functions: ParserNet and SizerNet use vertex, normal, Laplacian, and interpenetration losses, with ParserNet adding a weight regularizer for sparse local detail transfer.The ParserNet shape-under-clothing loss emphasizes visible skin and relaxes constraints over cloth regions.
- Implementation and experiments: Experiments hold out 40 scans per garment class, including unseen subjects and unseen sizes for known subjects, with batch size 8 and learning rate 0.0001.SizerNet uses latent dimension d = 30; networks are trained separately or fine-tuned by garment class as described.
5 Experiments and Results
Experiments evaluate ParserNet for garment parsing and SizerNet for size-conditioned garment resizing, using qualitative comparisons, error metrics, and baseline evaluations.
- 5.1 Results of 3D garment parsing and shape under clothing: ParserNet produces fine wrinkles and large garment deformations that a naive FC network cannot reproduce.Its localized sparse weighted reconstruction preserves input geometry details, whereas direct displacement prediction produces smooth fields.
- 5.2 Results of garment resizing: SizerNet evaluates resizing with surface-area error and per-vertex error, complemented by heatmaps because subtle shape changes are difficult to measure.Surface area captures garment scale, while per-vertex error captures folds and shape details associated with body shape and looseness or tightness.
- 5.1 Results of 3D garment parsing and shape under clothing: Table 1 reports average per-vertex parsing error across garment classes for the proposed method.The table measures Verr in millimeters for different garment classes.
- 5.2 Results of garment resizing: Figures 5–7 show ParserNet outputs and SizerNet resizing results for additional garments and smaller garment sizes.The visualizations include parsed multi-layer meshes, resized garments, and per-vertex error heatmaps.
- 5.2 Results of garment resizing: SizerNet yields lower resizing errors than linear scaling and average-garment baselines, indicating learned size variation.The evaluation compares error margins, average predictions, and linear scaling with alignment.
6 Conclusion
The paper introduces SIZER, ParserNet, and SizerNet for real-scan clothing parsing and size-sensitive garment modeling. The approach supports garment editing and resizing, but the dataset restricts the method to roughly aligned A-poses and geometry-only inputs.
- 6 Conclusion: SIZER captures clothing size variation across different subjects in a real dataset of roughly 2000 scans.The conclusion identifies SIZER as the first real dataset for this purpose.
- 6 Conclusion: ParserNet maps single mesh registrations to multi-layer garments and body shape under clothing without scan segmentation.SizerNet predicts garment draping across different sizes, enabling direct clothing-size changes on meshes.
- 6 Conclusion: The method is limited to roughly aligned A-poses and uses only geometry information for 3D clothing parsing.The authors plan to incorporate color through texture augmentation to improve accuracy and generalization.