Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

ScratchSim: A Procedural Synthetic Data Pipeline for Surface Scratch Detection

Source

arXiv

arxiv.org/abs/2607.27065v1

Read the source

Source

Plain-Language Brief

The paper introduces ScratchSim, a procedural rendering pipeline that generates large-scale synthetic data for surface scratch detection in industrial quality control. This pipeline uses BlenderProc and produces automatic annotations in COCO format. The researchers evaluate four training strategies and find that fine-tuning from synthetic weights consistently outperforms real-only training, and mixed training effectively recovers performance under scarce real-data conditions.

What the Document Says

  • The paper discusses the challenge of automated defect detection, specifically surface scratches in industrial quality control due to the scarcity of annotated defect data.
  • A procedural rendering pipeline is presented, which generates large-scale synthetic training data using BlenderProc. The pipeline allows for configurable material appearance, camera modes, and domain randomization, producing automatic COCO-format annotations.
  • Four training strategies are evaluated: synthetic-only, real-only, mixed, and fine-tuning from synthetic weights.
  • The evaluation is conducted across two objects with different material properties and three lightweight edge-deployable detectors: YOLOX, YOLO26, and LW-DETR.
  • The findings indicate that fine-tuning from synthetic weights consistently outperforms real-only training. Mixed training effectively recovers performance under scarce real-data conditions.
  • The researchers validate their findings across both convolutional and transformer-based architectures.
  • The approach enables scalable defect detection without the burden of large real annotated datasets, making it practical for on-device industrial inspection.
  • Upon acceptance, the pipeline scripts, 3D model, and both synthetic and real annotated scratch datasets for a glossy toy Ferrari car will be made available through the project website.

Key Concepts Mentioned

  • automated defect detection
  • surface scratches
  • industrial quality control
  • procedural rendering pipeline
  • BlenderProc
  • COCO format annotations
  • training strategies
  • synthetic-only
  • real-only
  • mixed
  • fine-tuning from synthetic weights
  • lightweight edge-deployable detectors
  • YOLOX
  • YOLO26
  • LW-DETR
  • convolutional and transformer-based architectures
  • scalable defect detection
  • large real annotated datasets

Claims Made by the Document

  • The scarcity of annotated defect data makes automated defect detection challenging.
  • ScratchSim enables scalable defect detection without the burden of large real annotated datasets.
  • Fine-tuning from synthetic weights consistently outperforms real-only training.
  • Mixed training effectively recovers performance under scarce real-data conditions.

Evidence Used

  • The text discusses the challenge of acquiring large amounts of annotated data for defect detection.
  • The text states that ScratchSim enables scalable defect detection without the burden of large real annotated datasets.
  • The text presents evaluation results showing that fine-tuning from synthetic weights outperforms real-only training.
  • The text provides evaluation results showing that mixed training effectively recovers performance under scarce real-data conditions.

Methods / Approach

  • procedural rendering pipeline
  • BlenderProc
  • configurable material appearance, camera modes, and domain randomization
  • automatic COCO-format annotations

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

The evaluation is conducted on a specific set of objects and detectors. The generalizability of the findings to other objects and detectors is unknown.

Open Questions

  • How well does ScratchSim perform on different types of surfaces or materials?
  • How well do the findings generalize to other object shapes and sizes?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The paper focuses on surface scratches on a specific type of object (a glossy toy Ferrari car) and evaluates performance with three specific detectors. It would be interesting to investigate how well the approach performs on different types of surfaces, materials, objects, and detectors.

Raw Abstract / Excerpt

While automated defect detection such as the detection of surface scratched is an important aspect in industrial quality control, the scarcity of annotated defect data make this task challenging. This paper presents a procedural rendering pipeline that generates large-scale annotated synthetic training data using BlenderProc, with configurable material appearance, camera modes, and domain randomization, producing automatic COCO-format annotations. To show the potential of our approach, we evaluate four training strategies, namely synthetic-only, real-only, mixed, and fine-tuning from synthetic weights, across two objects with different material properties and three lightweight edge-deployable detectors, YOLOX, YOLO26, and LW-DETR. Our evaluation show that fine-tuning from synthetic weights consistently outperforms real-only training, and that mixed training effectively recovers performance under scarce real-data conditions, with findings validated across both convolutional and transformer-based architectures. The proposed approach enables scalable defect detection without the burden of large real annotated datasets, making it practical for on-device industrial inspection. The pipeline scripts, 3D model, and both synthetic and real annotated scratch datasets for a glossy toy Ferrari car will be made available through the project website upon acceptance.