Fruits-3D — synthetic shape-biased source dataset for UDA

Fruits-3D is the synthetic source domain used in

Gemirter, C. B., Korkmaz, E. E., & Goularas, D. (2026). CoDA: A Cognitive-Inspired Approach for Domain Adaptation. Applied Sciences, 16(9), 4115. https://doi.org/10.3390/app16094115

It contains 12,000 rendered images (2,400 per class) of 5 fruit categories — apple, banana, pineapple, pomegranate, pumpkin — generated in Blender from 50 textured 3D assets (10 per class) under four rendering modes that form a shape-biased curriculum inspired by infant object individuation.

Rendering mode	Images / asset	Background	Purpose
HDR Nature	100	HDR maps of natural landscapes	"In-the-wild" clutter and non-uniform light
HDR Studio	50	HDR maps of indoor studios	Controlled, high-fidelity texture
Solid Color	50	Desaturated neutral solid	Foreground / background separation
Sculpture	40	Neutral solid, texture stripped (plaster-like)	Pure geometry — forces shape learning

A network trained from scratch on these 12,000 images alone (no ImageNet weights) and adapted with CoDA reaches an average Macro-F1 of 77.12 % across VegFru, Fruits-262 and Open Images v7 (5-class subsets), vs. 78.23 % for an ImageNet-pretrained baseline that saw 1.2 M real images (paper, Table 5).

Layout

data/
├── apple/
│   ├── hdr_nature_<model-md5>_<seq>.png
│   ├── hdr_studio_<model-md5>_<seq>.png
│   ├── solid_color_<model-md5>_<seq>.png
│   └── sculpture_<model-md5>_<seq>.png
├── banana/ …
├── pineapple/ …
├── pomegranate/ …
└── pumpkin/ …

One folder per class (so ImageFolder labels are the fruit classes); the rendering mode is the filename prefix and the 32-hex model-md5 identifies which of the 50 source 3D assets the view was rendered from (240 views per asset). Images are 576 × 384 PNG. This is exactly the folder the CoDA training code expects as storage.source_folder.

Loading with 🤗 Datasets:

from datasets import load_dataset
ds = load_dataset("cbalkig/Fruits-3D")                                   # label = fruit class
sculpture = load_dataset("cbalkig/Fruits-3D", data_files="data/*/sculpture_*.png")  # one mode only

Intended use

• Source domain for synthetic-to-real Unsupervised Domain Adaptation benchmarks (paper protocol: adapt to VegFru / Fruits-262 / Open Images v7 5-class subsets).
• Studying shape bias vs. texture bias: the Sculpture split is texture-free.
• Small-data / data-efficiency experiments (training from scratch on 12 k images).

How it was generated

The rendering pipeline (Blender, .usdz → 2D, pose / light / material / background randomisation) and the training code are in the CoDA repository: https://github.com/cbalkig/CoDA — module image_gen_3d_to_2d/.

The 50 source 3D assets come from SketchFab and are not redistributed here; their original licences apply. Only the rendered images are part of this dataset.

Citation

@Article{app16094115,
  author         = {Gemirter, Cavide Balkı and Korkmaz, Emin Erkan and Goularas, Dionysis},
  title          = {CoDA: A Cognitive-Inspired Approach for Domain Adaptation},
  journal        = {Applied Sciences},
  year           = {2026},
  volume         = {16},
  number         = {9},
  article-number = {4115},
  issn           = {2076-3417},
  doi            = {10.3390/app16094115},
  url            = {https://www.mdpi.com/2076-3417/16/9/4115}
}

Code archive: Zenodo DOI 10.5281/zenodo.23377177.