Reproduce “RoDLA: Benchmarking the Robustness of Document Layout Analysis Models” 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
Yufan Chen, Jiaming Zhang, Kunyu Peng, Junwei Zheng, Ruiping Liu, Philip Torr, Rainer Stiefelhagen
Researchers from Karlsruhe Institute of Technology and the University of Oxford established RoDLA, the first systematic robustness benchmark for Document Layout Analysis (DLA) models, encompassing 12 types of real-world image corruptions. They also developed a new robust DLA model, also named RoDLA, which achieved state-of-the-art performance with up to 7.1% higher P-Avg on perturbed datasets compared to previous methods, demonstrating improved resilience to image degradation.
Key insights
- Existing state-of-the-art DLA models show significant performance drops (e.g., up to 91.0% decrease for SwinDocSegmenter) even under moderate real-world image corruptions, highlighting a critical gap in practical deployability.
- The integration of dynamic spatial-wise average pooling and channel-wise attention mechanisms within the RoDLA architecture effectively extracts perturbation-insensitive features, leading to enhanced robustness without requiring specific pre-training on degraded data.
- One-stage robust DLA models, such as RoDLA, demonstrably outperform two-stage pipelines that include document image rectification, as rectification methods often fail or introduce additional distortions that hinder DLA performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.