metadata

dataset_info:
  features:
    - name: video
      dtype: string
    - name: question
      dtype: string
    - name: options
      list: string
    - name: answer
      dtype: string
    - name: answer_text
      dtype: string
    - name: meta
      dtype: string
    - name: source
      dtype: string
    - name: qa_subtype
      dtype: string
    - name: qa_type
      dtype: string
  splits:
    - name: test
      num_bytes: 515277
      num_examples: 1289
  download_size: 174366
  dataset_size: 515277
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
task_categories:
  - video-text-to-text

VideoEval-Pro

VideoEval-Pro is a robust and realistic long video understanding benchmark containing open-ended, short-answer QA problems. The dataset is constructed by reformatting questions from four existing long video understanding MCQ benchmarks: Video-MME, MLVU, LVBench, and LongVideoBench into free-form questions. The paper can be found here.

The evaluation code and scripts are available at: TIGER-AI-Lab/VideoEval-Pro

Dataset Structure

Each example in the dataset contains:

video: Name (path) of the video file
question: The question about the video content
options: Original options from the source benchmark
answer: The correct MCQ answer
answer_text: The correct free-form answer
meta: Additional metadata from the source benchmark
source: Source benchmark
qa_subtype: Question task subtype
qa_type: Question task type

Evaluation Steps

Download and Prepare Videos

# Navigate to videos directory
cd videos

# Merge all split tar.gz files into a single archive
cat videos_part_*.tar.gz > videos_merged.tar.gz

# Extract the merged archive
tar -xzf videos_merged.tar.gz

# [Optional] Clean up the split files and merged archive
rm videos_part_*.tar.gz videos_merged.tar.gz

# After extraction, you will get a directory containing all videos
# The path to this directory will be used as --video_root in evaluation
# For example: 'VideoEval-Pro/videos'

[Optional] Pre-extract Frames To improve efficiency, you can pre-extract frames from videos. The extracted frames should be organized as follows:

frames_root/
├── video_name_1/              # Directory name is thevideo name
│   ├── 000001.jpg             # Frame images
│   ├── 000002.jpg
│   └── ...
├── video_name_2/
│   ├── 000001.jpg
│   ├── 000002.jpg
│   └── ...
└── ...

After frame extraction, the path to the frames will be used as --frames_root. Set --using_frames True when running the evaluation script.

Setup Evaluation Environment

# Clone the repository from the GitHub repository
git clone https://github.com/TIGER-AI-Lab/VideoEval-Pro
cd VideoEval-Pro

# Create conda environment from requirements.txt (there are different requirements files for different models)
conda create -n videoevalpro --file requirements.txt
conda activate videoevalpro

Run Evaluation

cd VideoEval-Pro

# Set PYTHONPATH
export PYTHONPATH=.

# Run evaluation script with the following parameters:
# --video_root: Path to video files folder
# --frames_root: Path to video frames folder [For using_frames]
# --output_path: Path to save output results
# --using_frames: Whether to use pre-extracted frames
# --model_path: Path to model
# --device: Device to run inference on
# --num_frames: Number of frames to sample from video
# --max_retries: Maximum number of retries for failed inference
# --num_threads: Number of threads for parallel processing

python tools/*_chat.py \
    --video_root <path_to_videos> \
    --frames_root <path_to_frames> \
    --output_path <path_to_save_results> \
    --using_frames <True/False> \
    --model_path <model_name_or_path> \
    --device <device> \
    --num_frames <number_of_frames> \
    --max_retries <max_retries> \
    --num_threads <num_threads>

E.g.:
python tools/qwen_chat.py \
    --video_root ./videos \
    --frames_root ./frames \
    --output_path ./results/qwen_results.jsonl \
    --using_frames False \
    --model_path Qwen/Qwen2-VL-7B-Instruct \
    --device cuda \
    --num_frames 32 \
    --max_retries 10 \
    --num_threads 1

Judge the results

cd VideoEval-Pro

# Set PYTHONPATH
export PYTHONPATH=.

# Run judge script *gpt4o_judge.py* with the following parameters:
# --input_path: Path to save output results
# --output_path: Path to judged results
# --model_name: Version of the judge model
# --num_threads: Number of threads for parallel processing

python tools/gpt4o_judge.py \
    --input_path <path_to_saved_results> \
    --output_path <path_to_judged_results> \
    --model_name <model_version> \
    --num_threads <num_threads>

E.g.:
python tools/gpt4o_judge.py \
    --input_path ./results/qwen_results.jsonl \
    --output_path ./results/qwen_results_judged.jsonl \
    --model_name gpt-4o-2024-08-06 \
    --num_threads 1

Note: the released results are judged by gpt-4o-2024-08-06