FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents

By Guillaume Jaume, Hazim Kemal Ekenel, Jean-Philippe Thiran. 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), 2019

510 citations recorded by the source corpus.

We present a new dataset for form understanding in noisy scanned documents (FUNSD) that aims at extracting and structuring the textual content of forms. The dataset comprises 199 real, fully annotated, scanned forms. The documents are noisy and vary widely in appearance, making form understanding (FoUn) a challenging task. The proposed dataset can be used for various tasks, including text detection, optical character recognition, spatial layout analysis, and entity labeling/linking. To the best of our knowledge, this is the first publicly available dataset with comprehensive annotations to address FoUn task. We also present a set of baselines and introduce metrics to evaluate performance on the FUNSD dataset, which can be downloaded at https://guillaumejaume.github.io/FUNSD.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/58c793e278cdbf669a615b2c2479cd69ff785d63

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Chenliang Li
2. Songfang Huang
3. Ming Yan
4. Furu Wei
5. Bin Bi
6. Wei Wang
7. Shaohan Huang
8. Fei Huang

Citations in the corpus

• LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking (2022, methodology): “We evaluated pre-trained LayoutLMv3 models across five public benchmarks, including text-centric benchmarks: FUNSD [19] for form understanding, CORD [35] for receipt understanding, DocVQA [34] for document visual question answering, and image-centric benchmarks: RVL-CDIP [15] for document image clas”
• Summary of ChatGPT/GPT-4 Research and Perspective Towards the Future of Large Language Models (2023, methodology): “In the paper, ICL-D3IE is compared with existing pre-trained models such as LiLT,BROS (in-distribution (ID) setting and out-of-distribution (OOD) setting) on datasets (FUNSD, CORD, and SROIE [64, 65, 66]).”
• WebSRC: A Dataset for Web-Based Structural Reading Comprehension (2021, methodology): “Previous studies mainly focus on building templates for HTML DOM tree, called Wrapper Induction (Kushmerick, 2000; Flesca et al., 2004; Kushmer-ick et al., 1997; Muslea et al., 1999), or using well designed visual features like font sizes, element sizes, and positions (Zhu et al., 2005, 2006).”
• LVLM-EHub: A Comprehensive Evaluation Benchmark for Large Vision-Language Models (2023, unknown): “However, LVLMs that are pre-trained with relatively small pre-training datasets tend to struggle with image classification, particularly for fine-grained categories as depicted in Table VI, suffer from object hallucination as presented in Table IX.”
• StructuralLM: Structural Pre-training for Form Understanding (2021, background): “Table 1 presents the experimental results on the FUNSD test set.”
• Summary of ChatGPT-Related Research and Perspective Towards the Future of Large Language Models (2023, methodology): “88% on SROIE; in the out-of-distribution (OOD) setting, ICL-D3IE performs much better than previous pre-trained methods on all datasets.”
• OCRBench: on the hidden mystery of OCR in large multimodal models (2023, unknown): “We conduct experiments on SROIE [58], FUNSD [59] and POIE [60].”
• LayoutLM: Pre-training of Text and Layout for Document Image Understanding (2019, methodology): “for large-scale self-supervised pre-training. We select three benchmark datasets as the downstream tasks to evaluate the performance of the pre-trained LayoutLM model. The first is the FUNSD dataset3 [10] that is used for spatial layout analysis and form understanding. The second is the SROIE datase”
• VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding (2025, unknown): “…COCO-Text [66], TextOCR [67], BLIP3-OCR-Recap [58], LSVT [68], ReCTS [69] 4.69M Document SynthDoG-EN [70], SynthDoG-ZH [70], UReader-TR [71], FUNSD [72], DUDE [73], Vary-600k [74], pdfa-eng-wds Fine-grained Osprey-724K [76], MDVP-Data [77], ADE20K-Recap [78], Ob-ject365 [56], Flickr-30K [79],…”
• LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding (2020, methodology): “Entity Extraction Tasks Table 2 shows the model accuracy on the four datasets FUNSD, CORD, SROIE, and Kleister-NDA, which we regard as sequential labeling tasks evaluated using entity-level F1 score.”