SinerChen/SceneBench

Seeing the Scene Matters: a Scene-Aware Long-Video Benchmark

🤗 Benchmark | 📝 Paper

Overview

Long-video understanding remains challenging for multimodal large language models because real videos are not merely long sequences of frames, but are organized into semantically coherent scenes. Existing video benchmarks often emphasize short-clip perception or sparse frame matching, making it difficult to evaluate whether a model can understand scene-level events, connect multimodal cues across multi-minute videos, and reason over temporally distributed evidence.

SceneBench is a scene-aware long-video benchmark designed to evaluate long-video understanding at the scene level. It focuses on multi-minute videos and six task formats: Title Prediction, Comment Prediction, ClipQA, SceneQA, SceneQA-Audio, and I-VQA. After ambiguity filtering and quality control, the benchmark contains 8,507 final question-answer pairs.

This repository contains the complete SceneBench release archives and a five-record metadata split for the Hugging Face Dataset Viewer. The Viewer split covers SceneQA, ClipQA, iVQA, Comment, and Title, and exposes only video_id; it does not embed or render video media.

Highlights

• Scene-aware long-video benchmark: SceneBench targets long, multi-minute videos and emphasizes scene-level understanding rather than isolated frame perception.
• Manually annotated and cleaned data: All tasks are manually annotated, and ambiguous samples are removed during quality control.
• Multimodal scene-level reasoning: The benchmark evaluates the integration of visual, textual, and audio-related cues across coherent scene units.
• Practical long-video evaluation: The benchmark reflects realistic long-video settings where increasing the number of sampled frames may introduce irrelevant visual noise or exceed memory limits.

Tasks

task_type	Description
SceneQA	Scene-level questions requiring evidence from one or more temporal intervals.
SceneQA_Audio	Scene-level questions whose interpretation may depend on audio-related information.
ClipQA	Questions about a localized video clip.
iVQA	Inverse video question answering: select the question that matches the supplied answer.
Comment	Select the comment most likely associated with the video.
Title	Select the title most likely associated with the video.

The SceneQA Viewer record uses SceneQA_000632. The other four Viewer records were selected reproducibly with random seed 20260802.

Repository Structure

.
├── README.md
├── assets
│   └── stats.png
├── data
│   └── test.jsonl
├── evaluate.py
├── SceneBench.tar.part-000
├── ...
└── SceneBench.tar.part-061

The archive parts are retained as release artifacts and are not included in the Dataset Viewer configuration.

Evaluation

The included evaluate.py runs LongVA evaluation over the complete six-task SceneBench layout, supports resumable execution, decodes each video once for all associated questions, and writes detailed and per-task summary results.

Expected extracted data layout:

/path/to/SceneBench/
├── json
│   ├── sceneqa.json
│   ├── sceneqa_audio.json
│   ├── ClipQA.json
│   ├── ivqa.json
│   ├── comment.json
│   └── title.json
└── video
    ├── SceneQA
    ├── ClipQA
    ├── iVQA
    ├── comment
    └── title

Using LongVa as evaluation demo:

git clone https://github.com/EvolvingLMMs-Lab/LongVA.git
cd LongVA
conda create -n longva python=3.10 -y
conda activate longva
pip install --upgrade pip
pip install -e "longva[train]"

Run all six tasks with generic local paths:

export SCENEBENCH_ROOT=/path/to/SceneBench
export LONGVA_MODEL=/path/to/LongVA-7B

python evaluate.py \
  --data-dir "$SCENEBENCH_ROOT/json" \
  --video-root "$SCENEBENCH_ROOT/video" \
  --model-path "$LONGVA_MODEL" \
  --output-dir ./outputs \
  --run-name scenebench_longva_8frames \
  --device cuda:0 \
  --max-frames 8

Use --tasks to evaluate a subset, for example:

python evaluate.py \
  --data-dir "$SCENEBENCH_ROOT/json" \
  --video-root "$SCENEBENCH_ROOT/video" \
  --model-path "$LONGVA_MODEL" \
  --tasks SceneQA SceneQA_Audio

License

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

By accessing or using this dataset, users are expected to comply with the license terms. The dataset is provided strictly for non-commercial research use. Any use beyond this scope, including redistribution or application for commercial purposes, is not permitted, and users are responsible for any consequences resulting from improper use.

We do not claim ownership of the original video materials included in this dataset. The videos are provided only to support academic research, and all rights remain with their respective copyright holders. To reduce potential impact on the original works, the collected video clips have been processed through operations such as resolution reduction, temporal trimming, and format or size adjustment.

If a copyright holder believes that their content has been improperly included and wishes to request removal, contact sia1910023@gmail.com or submit an issue in this repository.

Citation

If you find this work useful for your research, please cite:

@inproceedings{anonymous2026seeing,
  title     = {Seeing the Scene Matters: a Scene-Aware Long-Video Benchmark},
  author    = {Sengnam Chen, Hao Chen, Chenglam Ho, Xinyu Mao, Jinping Wang, Yu Zhang, Chao Li},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026},
  url       = {https://arxiv.org/abs/2603.27259}
}