RoughCut-Bench: A Benchmark for Automatic Video Rough-Cut with Large Vision-Language Models|KDD 2027
Overview
RoughCut-Bench evaluates automatic video rough-cut in two settings:
- Visually driven: select highlights, order source media, and follow optional user requirements.
- Speech-driven: remove fillers, redundant expressions, and low-value content from ASR transcripts.
This dataset contains the refined face-blurred visual variant, professional visual annotations, and the speech-cleanup inputs and annotations. Media paths and visual ground-truth annotations are preserved from the prepared release. The speech component contains transcripts, not source talking-head videos.
Dataset
| Component | Cases | Source media | Highlight annotations |
|---|---|---|---|
| Visual Set-1: general rough-cut | 62 | 671 | 1,041 |
| Visual Set-2: with user requirements | 39 | 160 | 435 |
| Visual total | 101 | 831 | 1,476 |
| Speech-driven cleanup | 84 | — | — |
The distributed visual assets comprise 810 MP4 videos and 21 JPG images. The speech cases come from 20 source groups and contain 931 IM spans, 191 RM-to-R pairs, and 296 D spans under the companion evaluator's parser.
Download and use
Install the Hugging Face client, then download the repository while preserving its directory structure. The dataset is public and ungated; no login or access approval is required:
pip install -U huggingface_hub
hf download zhaors00/RoughCut-Bench \
--repo-type dataset \
--local-dir ./RoughCut-Bench-Data
export ROUGHCUT_DATA_ROOT="$(pwd)/RoughCut-Bench-Data"
Use this directory with the companion code. Generation and evaluation instructions are maintained in the code repository. This release preserves the benchmark's case-based file layout; use the download command above rather than loading the repository as a single flat table.
RoughCut-Bench-Data/
|-- benchmark_set1/<case_id>/<source_id>.mp4 # Also includes JPG assets
|-- benchmark_set2/<case_id>/<source_id>.mp4
|-- set1_label.json
|-- set2_label.json
|-- speech/<case_id>/
| |-- input.txt
| `-- gt.txt
|-- metadata/
| |-- summary.json
| |-- visual_manifest.jsonl
| |-- speech_manifest.jsonl
| `-- SHA256SUMS
`-- FACE_BLUR_PROCESSING.md
Annotation format
Visual. Each label JSON is an insertion-ordered mapping:
case_id -> source_id -> annotation record. Records contain highlights
(start/end times in seconds), confidence, duration, type, path, and
an optional prompt. Source-key order encodes the professional reference
ordering; do not alphabetically re-sort the label JSON. Resolve each path
relative to its corresponding benchmark_set1/ or benchmark_set2/ directory.
The duration annotation field is not the decoded media duration; consult
metadata/visual_manifest.jsonl for duration_seconds.
Speech. Each case provides an annotated gt.txt and an input.txt made by
removing only the annotation tokens. The input retains the original spoken
content; it is not a cleaned reference.
| Tag | Meaning |
|---|---|
[IM]... [IM] |
Filler or invalid marker to remove |
[RM]... [RM] |
Redundant or incorrect expression to remove |
[R]... [R] |
Corresponding expression to retain |
[D]... [D] |
Longer low-value span to remove |
Release processing and integrity
The visual assets use adaptive face-blur masks; original audio is retained.
Ten reviewed cat-only or product-only videos are byte-identical passthroughs.
See processing details. One explicit social handle
in the speech data is replaced by <REDACTED_SOCIAL_HANDLE> in both input and GT.
metadata/SHA256SUMS covers all 1,001 benchmark payload files: 831 visual
assets, two visual label files, and 168 speech files. To verify a download,
run the following from the downloaded dataset directory:
shasum -a 256 -c metadata/SHA256SUMS
License
The RoughCut-Bench authors' contributions are licensed under the Apache License 2.0. Third-party source material retains its original rights; this license grants only rights the contributors are authorized to grant.
- Downloads last month
- 9