<aside> 🔗 Link to this page: SciCap.AI

</aside>

<aside> 🔗 This year, the challenge will be hosted at IJCAI 2024 (August 3-9, Jeju Island, South Korea).

</aside>

1. Challenge Overview

Join the 2nd Scientific Figure Captioning (SciCap) Challenge! We will supply each team with approximately 400,000 scientific figure images from various arXiv papers, including their respective captions and relevant paragraphs. Teams will then use these data to build computational models to generate captions for these images. Whether you are working alone or as a team, we welcome researchers, AI/NLP/CV practitioners, and anyone interested in computational models for generating useful text for visuals to participate and submit their results.

This year, the challenge will be hosted at IJCAI 2024 (August 3-9, Jeju Island, South Korea).

Check out the details of the challenge, including data, code, baselines, evaluation criteria, and important dates. We eagerly await your participation in the 2nd SciCap Challenge!

For questions about the challenge, email us at [email protected].

Zoom Office Hour

The organizers will host a 30-minute-long Zoom office hour to answer all kinds of questions. Please do not hesitate to join us!

• Date: NOT Decided Yet
• Time: Eastern Daylight Time
• Zoom Link:

One Challenge, Two Track

The challenge will be structured into two tracks to accommodate the inherent differences and evaluation fairness between Short and Long captions. While longer captions tend to be viewed as more informative by readers [Huang et al. 2023], short captions are also crucial due to real-world space constraints in paper writing.

• Short Caption Track: To be eligible for the Short Caption Winner, a team must submit a result with 30% or more captions that are no longer than the original author-written captions per sample.
• Long Caption Track: To be eligible for the Long Caption Winner, a team must submit a result with 30% or more captions that are longer than the original author-written captions per sample.

Teams can submit multiple results during the submission period and choose which one to use for each track. Each team must participate in at least one track (The challenge will award two distinct winning teams. If one team tops both tracks, the runner-up in the Short Caption Winner track becomes the winner for that category). Teams will be allowed to use LLMs like GPT-4V and any external data for their captioning systems. However, using the original author-written captions from the test set will be prohibited.