Reproduce “CrashCar101: Procedural Generation for Damage Assessment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jens Parslov, Erik Riise, Dim P. Papadopoulos
A procedural generation pipeline, CrashCar101, was developed by researchers at the Technical University of Denmark to create a large synthetic dataset for vehicle damage assessment, featuring 101,050 images with pixel-accurate semantic part and damage annotations. Models pre-trained on CrashCar101 demonstrated mIoU improvements of +6.3-17.9% for 1-shot damage segmentation over general-purpose datasets, and consistently outperformed other synthetic data for part segmentation.
Key insights
- Procedural generation provides a scalable and cost-effective method to create extensive, pixel-accurate synthetic datasets, bypassing the limitations of manual annotation for complex semantic segmentation tasks.
- Carefully designed synthetic data with high diversity and realism, such as CrashCar101, can effectively reduce the sim2real domain gap for vehicle damage assessment, facilitating robust model pre-training.
- Training on a combination of synthetic and limited real-world data demonstrates synergistic benefits, often yielding superior performance compared to using either data source exclusively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.