CoM2-Bench: Grounding Confusable Movie Moments

Paper (arXiv 2610.23910) · Project page

Characters, locations, and actions recur throughout a movie, so many of its moments differ only in fine-grained details. CoM2-Bench evaluates temporal grounding in full-length movies with confusable query pairs: the two queries in each pair share three semantic attributes but differ in one, and they point to two temporally distant moments in the same movie. A model must ground each query in the full movie, so success requires resolving the single detail that tells the two moments apart.

This repository contains annotations only. Videos and subtitles are not redistributed (see Obtaining the videos).

Splits

Split	Movies	Hours	Pairs	Queries	Verification
test	46	90.4	1,507	3,014	Human-verified
train	64	123.5	2,865	5,730	Automatically generated, not human-reviewed

• The two splits use disjoint sets of movies.
• Use test for evaluation. train was used only to fine-tune a retriever in the paper and is released for reproducibility; it may contain noisy queries or intervals.
• Test statistics: average movie length 117.8 min; average query length 14.9 words; average moment length 60.2 s (median 42.5 s); paired moments are 35.3 min apart on average (median 29.7 min).

Data fields

Each line of data/{split}.jsonl is one query.

Field	Type	Description
movie	string	Movie title as it appears in ALLVB, e.g. "12 Years A Slave (2013)"
query_id	string	Query identifier, unique within a movie
pair_id	string	Pair identifier, unique within a movie
pair_role	string	"target" or "confusable"
paired_query_id	string	query_id of the other query in the pair
sentence	string	The natural-language query
timestamps	[float, float]	Ground-truth moment [start, end] in seconds of the full movie
movie_duration	float	Movie duration in seconds
shared_attributes	list[string]	The three attribute slots shared with the paired query
contrast_attributes	list[string]	The one attribute slot that differs from the paired query

Attribute slots: subject, core_action, motion_state, interaction_cause, context_setting, visual_anchor, temporal_composition.

query_id and pair_id are unique only within a movie, so use (movie, query_id) as the key.

Example:

{"movie": "12 Years A Slave (2013)", "query_id": "C003_U1_P001_T", "pair_id": "P001",
 "pair_role": "target", "paired_query_id": "C061_U1_P001_C",
 "sentence": "A man writes a letter using paper and ink in a dark room.",
 "timestamps": [184.225, 224.225], "movie_duration": 8050.167,
 "shared_attributes": ["subject", "core_action", "interaction_cause"],
 "contrast_attributes": ["context_setting"]}

Usage

from datasets import load_dataset

ds = load_dataset("queez0405/CoM2-Bench", split="test")

Obtaining the videos

The movies and subtitles come from ALLVB. Download them from ALLVB under its own terms of use, then match each annotation to its video by the movie title. The video and subtitle files are governed by ALLVB's terms, not by the license of this repository.

Evaluation

Each query is grounded independently in its full movie; the paired query, the pair role, and the attribute labels must not be given to the model as input. Missing or invalid predictions count as IoU 0.

• mIoU: mean temporal IoU over all queries.
• R@1@θ: percentage of queries whose top-1 prediction has IoU ≥ θ, for θ ∈ {0.3, 0.5, 0.7}.
• PairSuccess@θ: percentage of pairs for which both queries reach IoU ≥ θ.

Dataset creation

Queries were generated with Gemini 3.1 Pro in three stages: (1) subtitle-anchored clip segmentation, which groups consecutive subtitle cues into clips and describes each clip; (2) event-level attribute generation from context-expanded clip videos and descriptions, which proposes event intervals with structured attributes; and (3) confusable query pair generation, which builds parallel queries that share three attributes and contrast in one.

For the test split, every pair was reviewed by human annotators in a web interface that plays the movie with subtitles. Each query received a verdict in an initial pass and again in a second pass by a different annotator, and pairs with disagreeing verdicts were adjudicated. A pair was kept only if both queries were valid and unambiguous, the described content appeared within the ground-truth interval, and the interval tightly covered the event. See the appendix of the paper for details. The train split did not go through human verification.

License

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

Citation

@article{kim2026grounding,
  title   = {Grounding Confusable Movie Moments},
  author  = {Kim, Dong-Hee and Choi, Seonwoo and Kim, Changbeen and Ko, Juyeon and
             Choi, Youngju and Mun, Il Hyeon and Kim, Hyunwoo J. and Kim, Donghyun},
  journal = {arXiv preprint arXiv:2610.23910},
  year    = {2026}
}

Please also cite ALLVB if you use its videos.