Source-linked AI summary
Thingi10K: A Dataset of 10,000 3D-Printing Models
Qingnan Zhou, Alec Jacobson
TL;DR
3D-printing algorithms need benchmarks that reflect the varied and imperfect models encountered in fabrication, rather than sanitized visualization-oriented objects. The paper constructs and analyzes Thingi10K, a 10,000-mesh Thingiverse dataset with contextual annotations, geometric statistics, comparisons to existing datasets, and a query interface. The authors conclude that it provides a broad, balanced representation of real-world 3D-printing model complexity and quality, with the dataset and analysis publicly available.
Problem
3D-printing algorithms require test data representing real-world shape distributions, complexity, and fabrication-specific artifacts such as self-intersections and non-manifoldness.
Method
The authors systematically crawl Thingiverse to collect 10,000 models, annotate contextual information, analyze geometric properties, compare datasets, and provide a query interface.
Results
Thingi10K covers a large range of categories and provides a balanced representation of real-world 3D-printing data in mesh complexity and quality.
Takeaways & Limitations
The dataset and per-model analysis are freely available for stress-testing, learning algorithms, and exploring subsets by project-specific characteristics.
Takeaways & Limitations
The authors note that their mesh quality measures are incomplete and plan to expand the dataset over time as the 3D-printing community evolves.
Abstract
from arXiv · showhide
Empirically validating new 3D-printing related algorithms and implementations requires testing data representative of inputs encountered \emph{in the wild}. An ideal benchmarking dataset should not only draw from the same distribution of shapes people print in terms of class (e.g., toys, mechanisms, jewelry), representation type (e.g., triangle soup meshes) and complexity (e.g., number of facets), but should also capture problems and artifacts endemic to 3D printing models (e.g., self-intersections, non-manifoldness). We observe that the contextual and geometric characteristics of 3D printing models differ significantly from those used for computer graphics applications, not to mention standard models (e.g., Stanford bunny, Armadillo, Fertility). We present a new dataset of 10,000 models collected from an online 3D printing model-sharing database. Via analysis of both geometric (e.g., triangle aspect ratios, manifoldness) and contextual (e.g., licenses, tags, classes) characteristics, we demonstrate that this dataset represents a more concise summary of real-world models used for 3D printing compared to existing datasets. To facilitate future research endeavors, we also present an online query interface to select subsets of the dataset according to project-specific characteristics. The complete dataset and per-model statistical data are freely available to the public.
1 Introduction and background
Existing 3D-model datasets and standard test objects inadequately represent the geometric problems and contextual variety of models used for 3D printing. Thingi10K addresses this gap with 10,000 Thingiverse models, detailed analysis, comparisons, and public access tools.
- Standard models such as the Stanford bunny are smooth, manifold, and relatively simple, making them overly favorable test inputs.The paper argues that such models can provide a false sense of robustness when evaluating processing methods.
- Existing datasets are often sanitized or drawn from models intended for classification, scene understanding, or visualization rather than fabrication.The paper contrasts these datasets with the degeneracies, non-manifoldness, and self-intersections common in 3D-printing models.
- Thingi10K contains 10,000 Thingiverse models annotated with online tags and classes and analyzed for geometric properties including manifoldness, self-intersections, and genus.
- The dataset is intended for structural analysis, shape optimization, solid geometry, machine learning, and data-mining evaluation.
2 Methodology
The authors systematically crawled featured Thingiverse designs to construct Thingi10K, choosing a source with a large community, print-oriented models, and open licensing. The resulting collection contains 10,000 models from 2,011 things.
- Thingiverse was selected for its large active community, print-validated designs, and restriction to open-source licenses.
- Thingi10K consists of 10,000 models from 2,011 things collected systematically through web crawling.
- The sampling focuses on Thingiverse-featured things rather than random repository content, emphasizing designs selected for design, beauty, and manufacturability.
3 Analysis
Thingi10K is analyzed against MPZ14 and ShapeNetCore across model complexity, mesh quality, and contextual annotations. The results show that it spans real-world 3D-printing conditions, combining varied geometry with print-oriented metadata and quality issues.
- Dataset and measures: 10,000 models from 2,011 things are analyzed geometrically and contextually, with nearly all stored as STL files.The dataset also contains 42 OBJ, one PLY, and one OFF model.
- Complexity: Thingi10K covers a larger vertex-count range than MPZ14 and ShapeNetCore while retaining a similar distribution to ShapeNetCore.MPZ14 favors moderately high-resolution models and excludes extremely low- or high-resolution models.
- Complexity: 29% of Thingi10K models have multiple components, which typically represent designed parts, unlike many incoherent patches or isolated faces in ShapeNetCore.MPZ14 contains only single-component models, whereas ShapeNetCore is 83% multi-component.
- Complexity: Thingi10K’s genus distribution resembles MPZ14 but spans a wider range, with its highest genus exceeding MPZ14’s by more than 60 times.The comparison is restricted to single-component, closed, manifold meshes; no ShapeNetCore models satisfy those criteria.
- Mesh quality: MPZ14 is unrealistically pristine, while ShapeNetCore has mesh-quality problems not common to 3D-printed models.MPZ14 is perfect on seven quality measures, whereas ShapeNetCore performs poorly on six measures and has worse maximum and average triangle aspect ratios.
- Mesh quality: Thingi10K provides a broad range of mesh qualities, enabling stress tests by selected criteria; assuming self-intersection-free inputs would exclude 45% of models.The dataset includes both many high-quality models and a non-negligible share with common mesh-quality problems.
- Contextual information: Designer annotations include categories, subcategories, and 4,892 distinct free-form tags, which reveal differences in closure rates among modeling tools.98% of OpenSCAD models are closed, compared with 91% of SketchUp and 85% of TinkerCAD models.
- Contextual information: The fabrication-focused models often encode print-ready orientations and decompositions, and all things use open-source licenses.Their successful-print origin makes them ground-truth data for printing-oriented decisions such as orientation and decomposition.
4 Online query interface
The online query interface lets users filter Thingi10K models by contextual and geometric conditions, inspect matching results, and batch-download custom selections.
- Querying the dataset: Users specify one or more search clauses, and the interface retrieves models satisfying all conditions.An example clause is “genus>100”; clauses are comma-separated.
- Querying the dataset: Users can select single-component, manifold, solid meshes without self-intersections or degeneracies through a combined query.The interface lists every mesh satisfying the specified criteria.
- Querying the dataset: An auto-generated Python script supports batch downloading results for custom search terms.
- Model inspection: Model detail pages expose contextual and geometric information, identify authors and licenses, and link to the original Thingiverse pages.The interface is designed to respect each model’s copyright.
- Examples: Figure 14 demonstrates the interface with selected search results and their corresponding queries.
5 Conclusion
The paper presents Thingi10K, a freely available 10,000-mesh dataset of 3D-printing models with contextual and geometric analysis, plus a query interface for exploration. The authors position it as a practical resource for stress-testing and learning algorithms, while noting plans to expand it as the printing community evolves.
- Conclusion: Thingi10K contains 10,000 meshes systematically crawled from Thingiverse and analyzed using contextual and geometric information.The analysis compares the dataset with two existing 3D model datasets.
- Conclusion: The dataset covers many categories and offers a balanced representation of real-world mesh complexity and quality.
- Conclusion: The complete dataset and analysis are freely available, with a query interface for exploring and dissecting the collection.
- Conclusion: Thingi10K can serve as input for stress-testing and as ground truth for learning algorithms.
- Future work: The authors plan to update and enlarge the dataset to reflect the fast-evolving 3D-printing community.Planned additions include all featured Thingiverse things and user-suggested models.