CapRiCorn-1K: A Comprehensive Benchmark for Video Captioning and Subject Referential Consistency Across Temporal Scales
Paper • 2606.21949 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
CapRiCorn-1K is intended for research purposes only and must not be used for any commercial or other non-research purposes. Users assume full responsibility for any consequences arising from unauthorized use or redistribution.
Log in or Sign Up to review the conditions and access this dataset content.
CapRiCorn-1K: A Comprehensive Benchmark for Video Captioning and Subject Referential Consistency Across Temporal Scales
🎯 Overview
CapRiCorn-1K is the first benchmark dedicated to evaluating both video captioning quality and subject referential consistency across long temporal horizons and diverse scenarios.
Unlike existing benchmarks that mainly focus on short videos and overall caption quality, CapRiCorn-1K contains 1,000 manually collected videos spanning 15 seconds to 10 minutes, with rich scene transitions and recurring subjects. Besides evaluating caption accuracy and coverage, CapRiCorn-1K introduces a novel metric to assess whether the same subject is referred to consistently throughout the caption.
To support different captioning settings, CapRiCorn-1K provides:
- CapRiCorn-1K: audiovisual video caption evaluation (default).
- CapRiCorn-1K-V: visual-only video caption evaluation.
Extensive experiments show that current models struggle to maintain accurate descriptions and consistent subject references over long videos. Notably, the metrics on CapRiCorn-1K strongly correlate with downstream performance in both understanding and video generation tasks, validating its effectiveness as a proxy for real-world capabilities.
✨ Key Features
- 🎬 Long-Horizon Videos: 1,000 manually collected videos ranging from 15 s to 10 min (average 252 s).
- 🌍 Diverse Scenarios: Uniformly distributed across 8 major domains and 36 fine-grained subcategories.
- 🔄 Rich Scene Transitions: Every video contains at least one scene transition (average of 3.1 per video). Around 40% of videos involve subject clothing changes, making identity tracking highly challenging.
- 📝 Fine-Grained Annotation: Each video is manually labeled with an average of 4.4 subjects, 21.5 salient subject-related keypoints, and 14.9 other keypoints.
- 🔗 Strong Downstream Correlation: Evaluation metrics are proven to be highly correlated with downstream applications, accurately reflecting real-world utility.
- 🛡️ Robust Evaluation Protocol: The evaluation framework demonstrates high consistency and robustness across different judge models (e.g., closed-source GPT-4.1 and open-source Qwen3-235B-A22B-Instruct).
📏 Evaluation Metrics
CapRiCorn-1K utilizes an LLM judge (GPT-4.1 by default) to evaluate captions from two complementary perspectives based on manually annotated keypoints:
Overall Caption Quality
The judge model determines the mention status of every keypoint as: (a) Correctly mentioned; (b) Partially mentioned or mentioned with errors; (c) Not mentioned. We compute:
- Acc (Accuracy): Percentage of correctly mentioned keypoints.
- Cov (Coverage): Percentage of correctly or partially mentioned keypoints.
Subject Referential Consistency
For each mentioned subject, the judge model extracts all localized descriptions corresponding to that subject and performs co-reference clustering based on the caption context.
- Ref (Referential Consistency): Derived using a Rand Index-inspired approach. Measures how consistently the same subject is referred to throughout the caption. Higher Ref scores indicate better identity preservation and fewer ambiguous references.
🚀 Quick Start
Please refer to our Github repository for more details.
📜 License
CapRiCorn-1K is released under the CC-BY-NC-ND-4.0 license.
⚠️ Important Notice
CapRiCorn-1K is intended for research purposes only and must not be used for any commercial or other non-research purposes. Users assume full responsibility for any consequences arising from unauthorized use or redistribution.We do not claim copyright ownership of any raw video files. Video access is provided to researchers under the condition of compliance with the above license. We fully respect and acknowledge the copyrights of the original video creators.
If the original authors request removal of any videos, please contact us by email or submit an issue.
🖊️ Citation
If you find CapRiCorn-1K helpful for your research, please consider giving this repo a star ⭐ and citing our paper. We appreciate your support!
@article{chen2026capricorn,
title={CapRiCorn-1K: A Comprehensive Benchmark for Video Captioning and Subject Referential Consistency Across Temporal Scales},
author={Chen, Xinlong and Tang, Jiafu and Ding, Yue and Jia, Yizhuo and Li, Bozhou and Zeng, Bohan and Shi, Yang and Li, Shihao and Ji, Yiyan and Liu, Qiang and others},
journal={arXiv preprint arXiv:2606.21949},
year={2026}
}
- Downloads last month
- 30