LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding

Haoning Wu, Dongxu Li, Bei Chen, Junnan Li

Dataset on HuggingFace | Homepage | HuggingFace Leaderboard | ArXiv

Introduction

(left) An referring reasoning question. (right) Results with different input frames.

Initial Leaderboard

View more on HuggingFace Leaderboard.

[Custom Use] Load the LongVideoBench Dataset

1. Download the dataset via Hugging Face CLI:

huggingface-cli download longvideobench/LongVideoBench --repo-type dataset --local-dir LongVideoBench --local-dir-use-symlinks False

2. Extract from the .tar files:

cat videos.tar.part.* > videos.tar
tar -xvf videos.tar
tar -xvf subtitles.tar

3. Use the [LongVideoBench] dataloader to load the data from raw MP4 files and subtitles:

• (a) Install the dataloader:

git clone https://github.com/LongVideoBench/LongVideoBench.git
cd LongVideoBench
pip install -e .

• (b) Load the dataset in python scripts:

from longvideobench import LongVideoBenchDataset

# validation
dataset = LongVideoBenchDataset(YOUR_DATA_PATH, "lvb_val.json", max_num_frames=64)

# test
dataset = LongVideoBenchDataset(YOUR_DATA_PATH, "lvb_test_wo_gt.json", max_num_frames=64)

print(dataset[0]["inputs"]) # A list consisting of PIL.Image and strings.

The "inputs" are interleaved video frames and text subtitles, followed by questions and option prompts. You can then convert them to the format that your LMMs can accept.

[Automatic] Evaluating with LMMs-Eval

LongVideoBench has been integrated into LMMs-Eval library for automatic evaluation. With datasets and models on Hugging Face, you and can start automatic evaluation once the LMMs-Eval library is properly installed.

Install

Please install LMMs-Eval as follows:

git clone https://github.com/EvolvingLMMs-Lab/lmms-eval
cd lmms-eval
pip install -e .

This will install the GitHub main version that supports tasks: longvideobench_val_i (LongVideoBench for Image LMMs) and longvideobenc_val_v (LongVideoBench for Video-specific LMMs).

Example Use (Image LMMs)

We feed 16 frames by default for Image LMMs. To modify this, please go to lmms_eval/tasks/longvideobench/utils.py and change the parameter max_num_frames to other values (e.g. 4, 8, or 32, or even 64, 128, 256 for proprietary models).

• Idefics2

python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model idefics2 --tasks longvideobench_val_i --batch_size 1 --log_samples --log_samples_suffix idefics2_lvb_i --output_path ./logs/

• Phi3V

python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model phi3v --tasks longvideobench_val_i --batch_size 1 --log_samples --log_samples_suffix phi3v_lvb_i --output_path ./logs/

Example Use (Video-specific LMMs)

• LLaVA-NeXT-Video-34B-DPO

(32 frames)

python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model llavavid --model_args pretrained="lmms-lab/LLaVA-NeXT-Video-34B-DPO",max_frames_num=32,conv_template=chatml_direct,video_decode_backend="decord" --tasks longvideobench_val_v --batch_size 1 --log_samples --log_samples_suffix llavavid_34b_dpo_lvb_v --output_path ./logs/

• LLaVA-NeXT-Video-7B-DPO

(32 frames)

python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model llavavid --model_args pretrained="lmms-lab/LLaVA-NeXT-Video-7B-DPO",max_frames_num=32,video_decode_backend="decord" --tasks longvideobench_val_v --batch_size 1 --log_samples --log_samples_suffix llavavid_7b_dpo_lvb_v --output_path ./logs/

• Video-LLaVA

(8 frames)

python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model video_llava --tasks longvideobench_val_v --batch_size 1 --log_samples --log_samples_suffix video_llava_lvb_v --output_path ./logs/

Contact

Please contact haoning001@e.ntu.edu.sg for any queries.

License

This dataset follows CC-BY-NC-SA 4.0 license. Please use this dataset for non-commercial use ONLY.

Citation

@misc{wu2024longvideobench,
      title={LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding}, 
      author={Haoning Wu and Dongxu Li and Bei Chen and Junnan Li},
      year={2024},
      eprint={2407.15754},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.15754}, 
}