NarraScene

Paper (arXiv 2608.28699) · NeurIPS 2026 Evaluations & Datasets Track

NarraScene is a narrative-centric scene segmentation dataset for full-length movies. It accompanies Beyond Visual Boundaries: Rethinking Scene Segmentation for Movie RAG, which finds that existing scene segmentation methods fail to beat uniform chunking as retrieval units for movie RAG, because standard annotations emphasize visually salient transitions rather than narrative event structure.

Grounded in Event Segmentation Theory, every scene boundary in NarraScene requires a narrative-level shift, so the scenes capture coherent narrative events rather than visual cuts.

This repository contains annotations only. The movies come from MF² (see Obtaining the videos).

Boundary levels

Each scene boundary is labeled with one or more transition levels. The levels are not mutually exclusive.

Level	Meaning
1	Physical change in time, place, or other visually salient aspects of the scene
2	Change in the central character or entity configuration
3	Shift in the core goal, interaction, or conversational focus of the ongoing event

A valid narrative scene boundary must include Level 3, so every boundary in NarraScene contains Level 3, alone or with Level 1 and/or Level 2.

Levels	Scenes
1 + 2 + 3	1,266
1 + 3	489
2 + 3	403
3 only	213

Statistics

Movies	Hours	Mean movie length	Scenes	Mean scene length	Median scene length
53	78.0	88.3 min	2,371	118.4 s	99 s

Data fields

data/test.jsonl has one line per movie.

Field	Type	Description
movie_id	string	MF² movie id
file_name	string	MF² video file name, e.g. "1.mp4"
movie_duration	float	Duration of the MF² video in seconds
scenes	list	Scenes in temporal order
scenes[].scene_idx	int	1-based scene index
scenes[].start, scenes[].end	float	Scene start and end in seconds
scenes[].levels	list[int]	Transition levels of the boundary at the start of this scene

Scenes are contiguous from the start of each movie. In two movies the annotation stops slightly before or after the end of the video file: 48.mp4 leaves the last 2.8 minutes (end credits) unannotated, and the last scene of 50.mp4 ends 13.9 s after the video.

Example:

{"movie_id": "1", "file_name": "1.mp4", "movie_duration": 5630.163,
 "scenes": [{"scene_idx": 1, "start": 0, "end": 100, "levels": [3]}, ...]}

Usage

from datasets import load_dataset

ds = load_dataset("queez0405/NarraScene", split="test")

Obtaining the videos

The 53 movies are the open-licensed films of the MF² test set. Download them from sardinelab/MF2 under MF²'s terms, and match each annotation to its video by file_name (MF² stores them as test/<movie_id>.mp4). The videos and MF²'s own annotations are governed by MF²'s license, not by the license of this repository.

Annotation process

Four annotators labeled the movies with a custom web interface in three stages:

1. Independent proposal: one annotator watches the full movie, records boundary timestamps, and assigns levels.
2. Structured cross-verification: a second annotator reviews each candidate boundary and confirms, relabels, moves, deletes, inserts, or marks it as hard.
3. Consensus adjudication: all four annotators compare the original and revised annotations and resolve disagreements.

License

The annotations in this repository are released under CC BY 4.0. This license does not cover the movies, which must be obtained from MF² under its terms.

Citation

@inproceedings{kim2026beyond,
  title     = {Beyond Visual Boundaries: Rethinking Scene Segmentation for Movie RAG},
  author    = {Kim, Dong-Hee and Choi, Seonwoo and Kim, Changbeen and Wi, Jungmyung and Ko, Juyeon and
               Choi, Youngju and Mun, Il Hyeon and Kim, Hyunwoo J. and Kim, Donghyun},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}

Please also cite MF² if you use its videos.