Reproduce “synth-dacl: Does Synthetic Defect Data Enhance Segmentation Accuracy and Robustness for Real-World Bridge Inspections?” 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
Johannes Flotzinger, Fabian Deuser, Achref Jaziri, Heiko Neumann, Norbert Oswald, Visvanathan Ramesh, Thomas Braml
Researchers from German academic institutions introduced "synth-dacl," a set of synthetic dataset extensions, to enhance the accuracy and robustness of multi-class, multi-label semantic segmentation models for real-world concrete bridge defect inspection using the dacl10k dataset. This approach led to a 1.89% smaller drop in mean IoU and a 2.54% smaller drop in mean F1 score under various real-world perturbations, particularly improving fine-grained defect detection.
Key insights
- Targeted synthetic data effectively mitigates class imbalance and improves segmentation for challenging fine-grained defects like cracks and cavities.
- Synthetic data significantly enhances model robustness to diverse real-world image perturbations (e.g., noise, blur, illumination changes, weather effects), leading to less performance degradation.
- Fine-resolution synthetic annotations, especially for cracks, enable models to predict defects with greater pixel-level accuracy than training solely on coarse real-world labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.