Reproduce “CarDD: A New Dataset for Vision-based Car Damage Detection” 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
Xinkuang Wang, Wenjing Li, Zhongcheng Wu
A new large-scale, high-resolution dataset named CarDD is introduced to advance vision-based car damage detection and segmentation, providing over 9,000 instances across six damage types. Experiments using state-of-the-art models and a proposed DCN+ demonstrate improved performance, particularly for subtle damages like dents, scratches, and cracks, and highlight the potential of Salient Object Detection for irregular shapes.
Key insights
- High-resolution images and fine-grained pixel-level annotations, as provided by CarDD, are crucial for training models capable of detecting subtle and small car damages.
- Integrating focal loss and multi-scale learning into detection models like DCN+ significantly improves performance for challenging damage types (dent, scratch, crack) and small-scale objects.
- Salient Object Detection (SOD) methods show promise in accurately segmenting irregular and slender damage shapes, suggesting their potential to complement instance segmentation for detailed damage localization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.