FALCONEye: Finding Answers and Localizing Content in ONE-hour-long videos with multi-modal LLMs

This repo contains the benchmark presented in the paper "FALCONEye: Finding Answers and Localizing Content in ONE-hour-long videos with multi-modal LLMs".

🔔 News:

🥳 3/26/2025: We have released the FALCON-Bench and Paper! 🔥

Introduction

Information retrieval in hour-long videos presents a significant challenge, even for state-of-the-art Vision-Language Models (VLMs), particularly when the desired information is localized within a small subset of frames. However, few public benchmarks are available to assess this challenge. To address this, we introduce FALCON-Bench, a question-answering benchmark with one-hour-long videos that focuses on open-ended question evaluation and contains ground truth temporal information about when the answer is located.

The main challenge that FALCON-Bench targets is to accurately pinpoint the answer and the corresponding short temporal clip that contains the answer within one-hour-long videos. We refer to this task as Video Answer Search.

The test set comprises 506 questions built over 80 videos. Regarding evaluation, most VLMs that are limited to a small number of sampled frames fail to outperform the LLM-blind baselines in multiple-choice evaulation mode. This performance slightly improves when models are prepared for long video understanding. However, only FALCONEye-a novel video agent which combines a VLM and a LLM guided by an exploration algorithm- is able to reason until properly tackling the task. These results establish FALCON-Bench as a valuable benchmark for future long-context LMMs and video agents.

Download videos

Videos have been sourced from three different recognized datasets as: Soccernet, MovieChat-1K and Walking_Tours. While Walking_Tours videos are already in the repo, we provide the code to download the remaining videos from both SoccerNet and MovieChat-1K datasets, respectively.

Fill this Non Disclosure Agreement form of SoccerNet dataset and save the password (soccernet_pwd) that they will send to you by mail.
Request access to MovieChat-1K dataset.
Install FALCON-Bench hf repo. Select the path hf_path where it will be downloaded (45GB).

huggingface-cli download cplou99/FALCON-Bench --repo-type dataset --local-dir hf_path
cd hf_path

Create conda environment with SoccerNet pip package.

conda env create -f FALCON-Bench-env.yml
conda activate FALCON-Bench

Run the download_videos.py script to download and organize data. Fill the arguments: SoccerNet password, FALCON-Bench hf repo path, and the desired downloaded data_path (155GB).

python download_videos.py --soccernet_password <soccernet_pwd> --hf_benchmark_dir "./" --data_dir <data_path>

As result, you will get 81 videos inside a folder named "full_videos".

FALCON-Bench

506 Q&As built over 80 one-hour-long videos sourced from three different datasets. In addition to the answer, we provide the ground truth temporal clip (gt_time_interval) and a ground truth frame (gt_frame_idx) within the answer is contained.

License

Our dataset is under the CC-BY-NC-SA-4.0 license.

⚠️ If you need to access and use our dataset, you must understand and agree: This dataset is for research purposes only and cannot be used for any commercial or other purposes. The user assumes all effects arising from any other use and dissemination.

We do not own the copyright of any raw video files. Currently, we provide video access to researchers under the condition of acknowledging the above license. For the video data used, we respect and acknowledge any copyrights of the video authors. Therefore, for the movies, TV series, documentaries, and cartoons used in the dataset, we have reduced the resolution, clipped the length, adjusted dimensions, etc. of the original videos to minimize the impact on the rights of the original works.

If the original authors of the related works still believe that the videos should be removed, please contact c.plou@unizar.es or directly raise an issue.

Citation

If you find this repository useful, please consider giving a star 🌟 and citation

@article{plou2025falconeyefindinganswerslocalizing,
      title={FALCONEye: Finding Answers and Localizing Content in ONE-hour-long videos with multi-modal LLMs}, 
      author={Carlos Plou and Cesar Borja and Ruben Martinez-Cantin and Ana C. Murillo},
      year={2025},
      eprint={2503.19850},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.19850},
}

Datasets:

cplou99
/

FALCON-Bench

You need to agree to share your contact information to access this dataset