ThingiPrint

ThingiPrint is a benchmark dataset that links CAD models to real photographs of their 3D-printed counterparts.
It is intended as an evaluation benchmark for 3D-printed object classification, focusing on generalization to previously unseen object instances.

Associated paper

Classifying Novel 3D-Printed Objects without Retraining: Towards Post-Production Automation in Additive Manufacturing

Object overview

Dataset contents

Folder names and labels are the original Thingi10K file IDs. The same IDs are used consistently for the real photographs, synthetic reference renders, and source meshes.

Main dataset (main)

• 100 object classes
• 1,000 real photographs
• Approximately 10 photographs per object
• Industrial SLS printer: Sindoh S100
• Material: white PA12 (polyamide 12)

Auxiliary dataset (auxiliary)

Twenty objects were reprinted to evaluate robustness to a different printer and material.

• 20 object classes
• 200 real photographs
• Approximately 10 photographs per object
• Printer: Prusa MK4
• Material: white PLA

Synthetic reference renders (renders)

These are the exact synthetic images used to construct the class prototypes for the main evaluation reported in the paper.

• 100 object classes matching the main dataset
• 24 renders per object (2,400 PNG images total)
• 256 x 256 pixels
• Two elevation rings, with 12 azimuths per ring

Source meshes

ThingiPrint_CAD/meshes contains the 100 Thingi10K STL meshes corresponding to the main benchmark objects. ThingiPrint_CAD/object_metadata.csv records the Thingi10K file ID, Thingiverse thing ID, creator, title, original license, license URL, source URLs, and SHA-256 checksum for every mesh.

Loading the image configurations

from datasets import load_dataset

real = load_dataset(
    "fanismathioulakis/thingiprint", "main", split="test"
)
auxiliary = load_dataset(
    "fanismathioulakis/thingiprint", "auxiliary", split="test"
)
renders = load_dataset(
    "fanismathioulakis/thingiprint", "renders", split="test"
)

Each image example has an image field and a class-label field. The label names are the corresponding Thingi10K file IDs.

Downloading the source meshes

Install the Hugging Face CLI and download all 100 meshes plus their metadata:

pip install -U huggingface_hub

hf download fanismathioulakis/thingiprint \
  --repo-type dataset \
  --include "ThingiPrint_CAD/**" \
  --local-dir thingiprint-cad

To download a single object by its Thingi10K file ID:

hf download fanismathioulakis/thingiprint \
  ThingiPrint_CAD/meshes/34785.stl \
  --repo-type dataset \
  --local-dir thingiprint-cad

Data organization

thingiprint/
|-- ThingiPrint/
|   `-- test/<Thingi10K file ID>/*.jpg
|-- ThingiPrint_Auxiliary/
|   `-- test/<Thingi10K file ID>/*.jpg
|-- ThingiPrint_Renders/
|   `-- test/<Thingi10K file ID>/*.png
`-- ThingiPrint_CAD/
    |-- meshes/<Thingi10K file ID>.stl
    |-- object_metadata.csv
    |-- SHA256SUMS
    `-- README.md

Licenses

This repository is mixed-license:

• Real photographs: CC BY 4.0.
• Source meshes (ThingiPrint_CAD/meshes): each mesh keeps its original Thingi10K/Thingiverse license (CC BY, CC BY-SA, CC BY-NC, CC BY-NC-SA, GPL, or public domain), as listed in ThingiPrint_CAD/object_metadata.csv.
• Synthetic renders (ThingiPrint_Renders): derived from the meshes, each render folder follows the license of the mesh with the same Thingi10K file ID.

Check object_metadata.csv before using or redistributing a mesh or its renders. Some licenses require attribution, share-alike, or non-commercial use.

Thingi10K is curated by Qingnan Zhou and Alec Jacobson: https://github.com/Thingi10K/Thingi10K

Citation

@article{mathioulakis2026classifying,
  title={Classifying Novel 3D-Printed Objects without Retraining: Towards Post-Production Automation in Additive Manufacturing},
  author={Mathioulakis, Fanis and Radevski, Gorjan and Cleuren, Silke GC and Janssens, Michel and Das, Brecht and Schauwaert, Koen and Tuytelaars, Tinne},
  journal={arXiv preprint arXiv:2603.07465},
  year={2026}
}