hminjeong/TripleSumm-MoSu

Dataset Summary

MoSu (Most Replayed Multimodal Video Summarization) is the first large-scale multimodal video summarization dataset. It provides synchronized visual, audio, and text features for 52,678 in-the-wild videos. The ground-truth annotations are based on YouTube's "Most Replayed" statistics, offering highly reliable per-frame importance scores derived from collective viewer engagement.

• Paper: TripleSumm: Adaptive Triple-Modality Fusion for Video Summarization
• GitHub Repository: smkim37/TripleSumm

Dataset Structure

The dataset consists of 6 core files providing metadata, multimodal features, ground truth annotations, and evaluation splits.

1. Metadata (mosu_metadata.csv)

Contains the foundational information for all 52,678 videos.

• video_id: The unique identifier for the video. This serves as the universal key to access data in all .h5 files and the split JSON.
• youtube_id: The original YouTube video ID. The video can be accessed via https://www.youtube.com/watch?v={youtube_id}.
• duration: The length of the video in seconds.
• views: The total view count of the video.
• labels: Original multi-label annotations provided by the YouTube-8M dataset.
• cluster_id: One of 10 semantic clusters (0-9). These clusters were generated based on metadata to group videos by topic (e.g., Video Games, Sports) and ensure a balanced distribution across dataset splits. For more details, please refer to the original paper.

2. Multimodal Features (.h5 files)

Pre-extracted features for all three modalities. Each file is provided in HDF5 format and is approximately 40GB in size. All features have a shape of (N, D), where N corresponds to the video's duration (in seconds) indicated in the metadata, and D is 768 for all modalities.

• mosu_feat_visual_clip.h5: Visual features extracted using CLIP.
• mosu_feat_audio_ast.h5: Audio features extracted using Audio Spectrogram Transformer (AST).
• mosu_feat_text_roberta.h5: Text features extracted using RoBERTa.

File Size & Downloading: The feature files are extremely large (~40GB each). Depending on network conditions, downloading may take a considerable amount of time.

3. Ground Truth (mosu_gt.h5)

An HDF5 file containing the summarization labels for all 52,678 videos. Each video_id (e.g., '005O') maps to an HDF5 Group containing four specific keys:

• change_points: Temporal boundaries for video shots.
• cluster_id: The semantic cluster ID of the video.
• gt_score: Frame-level ground-truth importance scores.
• gt_summary: Binary labels indicating whether a frame is included in the final summary.

4. Dataset Splits (mosu_split.json)

Contains standardized splits for training, validation, and testing. The split ratio strictly maintains the proportional representation of each cluster_id for a balanced evaluation.

• train_keys: List of video IDs for 42,152 training videos.
• val_keys: List of video IDs for 5,263 validation videos.
• test_keys: List of video IDs for 5,263 testing videos.

Citation

If you use the MoSu dataset or the TripleSumm model in your research, please cite the following paper:

@inproceedings{triplesumm2026,
  title={TripleSumm: Adaptive Triple-Modality Fusion for Video Summarization},
  author={Kim, Sumin and Jeong, Hyemin and Kang, Mingu and Kim, Yejin and Oh, Yoori and Lee, Joonseok},
  booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
  year={2026}
}