saptarshineilsinha/ToyFerrariCarScratchDefectDataset

Datasets — Toy Ferrari Car

All datasets below belong to the toy Ferrari car dataset and are organised by detector (LW-DETR, YOLO26) and by training regime (synthetic only, real only, infused = synthetic + scarce real).

Overview

Detector	Regime	Background	Path
LW-DETR	Infused (synth + scarce real)	random	lwdetr/infusion_rb
LW-DETR	Infused (synth + scarce real)	white (WB)	lwdetr/infusion_wb
LW-DETR	Synthetic only	random	lwdetr/synth/synth_lwdetr_random_background.zip
LW-DETR	Synthetic only	white (WB)	lwdetr/synth/synth_lwdetr_white_background.zip
LW-DETR	Real only (scarce subsets)	real	lwdetr/real_scarce
LW-DETR	Real only (100 %)	real	lwdetr/real_scarce/synth_mixed_scarce_lwdetr_real100
YOLO26	Infused (synth + scarce real)	random	yolo26/synth_mixed_scarce_random_background.zip
YOLO26	Infused (synth + scarce real)	white (WB)	yolo26/synth_scarce_white_background.zip
YOLO26	Synthetic only	random	yolo26/synth_random_background.zip
YOLO26	Synthetic only	white (WB)	yolo26/synth_white_background.zip
YOLO26	Real only	real	yolo26/real_data_scratches.zip

Directory layout

lwdetr/
├── infusion_rb/                                   # infused, random background
├── infusion_wb/                                   # infused, white background
├── real_scarce/                                   # real only (10 % / 25 % / 50 %)
│   └── synth_mixed_scarce_lwdetr_real100/          # real only (100 %)
└── synth/
    ├── synth_lwdetr_random_background.zip          # synthetic only, random bg
    └── synth_lwdetr_white_background.zip           # synthetic only, white bg

yolo26/
├── synth_mixed_scarce_random_background.zip        # infused, random background
├── synth_scarce_white_background.zip               # infused, white background
├── synth_random_background.zip                     # synthetic only, random bg
├── synth_white_background.zip                      # synthetic only, white bg
└── real_data_scratches.zip                         # real only

Notes

• Background modes: random background (default) renders synthetic images on randomised backgrounds; white background (WB) mimics the static white backdrop of the real capture setup.
• Infused datasets contain synthetic images mixed with a scarce fraction of real images (10 %, 25 %, 50 %, 100 %); the fraction is selected via the respective subset folder / config.
• Evaluation is always performed on a held-out test set of real images that is never used during training.
• Fine-tuning regime (Synth → Real) uses the synthetic only datasets for pre-training and the real only dataset for the second stage.
• LW-DETR expects COCO-style annotations, YOLO26 the Ultralytics YOLO format (images/, labels/, data.yaml).

Citation

If you use this dataset, please cite our paper:

@misc{kühn2026scratchsimproceduralsyntheticdata,
      title={ScratchSim: A Procedural Synthetic Data Pipeline for Surface Scratch Detection}, 
      author={Paul Julius Kühn and Saptarshi Neil Sinha and Tiago Kleist and Richard Hoffmann and Arjan Kuijper and Michael Weinmann},
      year={2026},
      eprint={2607.27065},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.27065}, 
}

License

Released under CC BY 4.0.