Reproduce “InfoDet: A Dataset for Infographic Element 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
Jiangning Zhu, Yuxing Zhou, Zheng Wang, Juntao Yao, Yima Gu, Yuhui Yuan, Shixia Liu
InfoDet provides a grounding method that helps vision-language models connect questions to elements in infographics. Detection and OCR identify charts, human-recognizable objects, and text; labeled boxes and descriptions guide the model’s reasoning. The dataset contains 101,264 infographics. On ChartQAPro, o4-mini reaches 64.9% enhanced relaxed accuracy with grounded chain-of-thought, versus 63.2% with direct prompting, on this benchmark and metric.
Key insights
- Explicitly grounding infographic elements (texts, charts, HROs) and using a "Thinking-with-Boxes" scheme with grounded chain-of-thought reasoning significantly improves VLM performance on complex infographic understanding tasks.
- Existing state-of-the-art object detection models show limited zero-shot and few-shot performance on infographic elements, highlighting the necessity of domain-specific pre-training or fine-tuning with a dedicated dataset like InfoDet.
- Models fine-tuned on InfoDet exhibit stronger generalizability to other graphic layout detection tasks, suggesting that learning infographic structures can benefit broader visual document analysis, such as UI component and document layout detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.