Reproduce “Improving Table Structure Recognition with Visual-Alignment Sequential Coordinate Modeling” 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
Yongshuai Huang, Ning Lu, Dapeng Chen, Yibo Li, Zecheng Xie, Shenggao Zhu, Liangcai Gao, Wei Peng
The VAST framework enhances table structure recognition by employing a coordinate sequence decoder for precise bounding box prediction and a visual-alignment loss that enriches logical representations with local visual information. This end-to-end approach achieved state-of-the-art results on several benchmarks, including 98.63% S-TEDS on FinTabNet and 97.23% S-TEDS on PubTabNet.
Key insights
- Modeling bounding box coordinates as a language sequence and predicting them sequentially leverages inter-coordinate dependencies more effectively than direct regression for enhanced spatial localization.
- Integrating a visual-alignment loss during training compels logical structure representations to incorporate precise local visual features, which is crucial for accurate bounding box prediction.
- By ensuring the HTML decoder's attention focuses on local visual features of non-empty cells, the visual-alignment loss indirectly supports better bounding box generation by the coordinate decoder.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.