OmniVDCBench: Caption-as-Evidence Evaluation of Audio-Visual Dense Captioning

Dataset Card · Audio-Visual Dense Captioning · Caption-as-Evidence Evaluation · Video Question Answering

⚠️ Note on video files. This repository contains only the annotation files (three JSON files), the benchmark figure, and this dataset card. The 1,253 raw video files referenced by the annotations are not uploaded here. The raw video files will be released after the associated paper is accepted.

Figure 1. Illustration of the OmniVDCBench caption-as-evidence evaluation setting.

---

1. Overview

OmniVDCBench is a benchmark for evaluating audio-visual dense captioning under a caption-as-evidence paradigm. Instead of asking models to answer questions directly from raw video, OmniVDCBench anchors each question-answer (QA) pair in a verified fused caption and asks the evaluated model to produce a description that preserves sufficient evidence to support a correct answer.

Recent multimodal large language models (MLLMs) have substantially improved dense video captioning, yet existing benchmarks rarely assess whether the generated caption itself retains the audio and visual evidence needed for downstream reasoning. OmniVDCBench addresses this gap by evaluating captions through evidence-grounded, multiple-choice QA.

The benchmark is constructed with the Audio-Visual Value Diagnostic (AVVD) framework, which decouples modality diagnosis from caption fusion to reduce redundancy and cross-modal conflicts. Candidate QA pairs are generated by DeepSeek-V3.2 across four dimensions and verified by human annotators.

OmniVDCBench comprises 1,253 videos across 8 scenarios and 4,794 evaluation items, supported by more than 6,000 annotator-hours. Across 15 representative models, the best proprietary and open-weight systems achieve 74.36% and 55.96% Overall Accuracy, respectively, while a separate expert evidence-writing study on the intangible cultural heritage (ICH) subset reaches 98.20% accuracy.

---

2. Audio-Visual Value Diagnostic (AVVD) Framework

AVVD mitigates information loss, redundancy, and cross-modal conflicts from naive concatenation through two stages.

1. Semantic value diagnosis. The audio transcript and visual caption are split into semantically aligned segment pairs (approximately three segments per minute, without rigid timestamp alignment). A diagnostic model analyzes each pair and assigns the non-dominant modality one of four roles: complementary, explanatory, overlapping, or weak.
2. Report-guided caption fusion. A fusion model synthesizes the aligned segments into a unified caption by following the diagnosis report: it preserves dominant-modality information, incorporates complementary and explanatory evidence, merges overlapping content, and removes weak or redundant information. Human annotators then correct any missing, unsupported, or inconsistent content.

In practice, the construction pipeline first separates audio and visual streams with FFmpeg; Qwen3-ASR-1.7B produces the audio transcript, Qwen3-VL-32B-Instruct produces the visual caption, and AVVD fuses them into the reference audio-visual caption.

---

3. Question Dimensions

Caption-anchored items are constructed across four dimensions:

Dimension	Abbrev.	Focus
Independent events	IE	Salient events that can be interpreted locally
Associated events	AE	Temporal, causal, or semantic relations across events
Hallucinated events	HE	Correct options paired with plausible but nonexistent-event distractors
Fine-grained details	FD	Specific visual, auditory, spatial, and temporal information

The released benchmark contains 1,563 independent-event, 727 associated-event, 587 hallucinated-event, and 1,917 fine-grained-detail items (4,794 in total).

---

4. Benchmark Composition

4.1 Data Collection and Distribution

OmniVDCBench contains 1,253 unique videos and 4,794 evaluation items. Videos are sampled from WanJuan1.0, LLaVA-Video-178K, and InternVid-10M-FLT, together with user-generated videos from Bilibili and TikTok and intangible cultural heritage (ICH) videos provided by the National Library of China.

The 8 scenarios are daily activities, outdoor scenes, sports, handicrafts, humanities, ICH crafts, ICH Chinese cuisine, and ICH cultural customs. Durations range from short clips to over 5 minutes, and resolutions from 360p to at least 1080p.

4.2 Human Annotation and Quality Check

Nine annotators contributed over 6,000 annotator-hours across three months reviewing videos, captions, QA pairs, and evidence. Multiple review rounds checked option exclusivity and plausibility, diagnostic value, dimension consistency, and explicit evidence support. Of the candidate items, 9.2% were retained unchanged, 70.7% were revised, and 20.1% were discarded.

---

5. Released Files and Splits

The repository releases annotation files only. Each JSON file is a list of QA objects; every object records the video filename, video description, video category, question type, question text, answer options, gold answer, and reference evidence.

Split	Annotation File	Videos	QA Records	Language	Video Filename Pattern
heritage_zh	非遗视频类型_1243.json	330	1,244	Chinese	0001.mp4 … 0330.mp4
general_zh	通用视频类型_2343.json	523	2,343	Chinese	0331.mp4 … 0853.mp4
general_en	通用视频类型_英文_1205.json	400	1,207	English	0854.mp4 … 1253.mp4
Total	—	1,253	4,794	Chinese + English	—

Video Category Distribution

heritage_zh (330 videos)

video_type	Videos	QA Records
ICH-Craft	154	583
ICH-Cuisine	151	566
ICH-Custom	25	95

general_zh (523 videos)

video_type	Videos	QA Records
Daily	258	1,103
Outdoor	105	458
Humanities	59	270
Handcrafts	52	276
Sports	49	236

general_en (400 videos)

video_type	Videos	QA Records
Daily	100	305
Outdoor	100	302
Handcrafts	100	301
Sports	100	299

The video_name field references the raw video file. Raw videos are not distributed in this repository; see Section 8 for the release policy.

---

6. Data Format

Each JSON file is a list of flattened QA records. A simplified example is shown below:

{
  "id": "通用视频类型_2343_000001_001",
  "source_file": "通用视频类型_2343.jsonl",
  "source_line": 1,
  "video_name": "0001.mp4",
  "video_type": "Handcrafts",
  "video_description": "A detailed natural-language description of the video content...",
  "question_type": "independent-event",
  "question_text": "Question text...",
  "options": {
    "A": "Option A",
    "B": "Option B",
    "C": "Option C",
    "D": "Option D"
  },
  "correct_answer": "B",
  "reference_text": "Supporting evidence or reference span..."
}

Field Description

Field	Type	Description
id	string	Unique QA record identifier.
source_file	string	Name of the intermediate source file from which the QA record was produced.
source_line	integer	Source line index in the intermediate annotation file.
video_name	string	Video filename (0001.mp4 … 1253.mp4, continuous across the three splits). The corresponding raw video file is released separately.
video_type	string	Coarse-grained video category (one of the 8 scenarios).
video_description	string	Long-form description of the video content.
question_type	string	Question dimension label (independent-event, associated-event, hallucinated-event, or fine-grained-detail).
question_text	string	Question prompt.
options	object	Candidate answer options.
correct_answer	string	Gold answer label, e.g. A, B, C, D, or a multi-label answer such as AB.
reference_text	string	Supporting evidence text used for answer verification.

---

7. Evaluation Protocol

Under the caption-as-evidence protocol, the evaluated model generates an audio-visual caption. A judge model receives only the generated caption, the question, and the answer options — without access to the original media or reference caption — and predicts an answer together with the supporting text it relies on.

The judge is instructed to return UNKNOWN when the caption lacks sufficient evidence to uniquely support an answer (including missing or conflicting evidence). Reported metrics are:

• Answer Rate — the proportion of non-UNKNOWN predictions (1 − UNKNOWN rate).
• Conditional Accuracy — accuracy among non-UNKNOWN predictions.
• Overall Accuracy — accuracy over all questions, with UNKNOWN counted as incorrect.

Qwen3.5-27B serves as the judge in all main experiments. Evaluations with eight judge models yield pairwise Spearman correlations of 0.939–0.997, indicating highly consistent model rankings across judge configurations.

---

8. Video Release Policy

The 1,253 raw video files are not hosted in this repository. The videos may be subject to copyright, platform, or redistribution constraints, and are distributed separately.

The raw video files will be released after the associated paper is accepted. Please do not redistribute the video files without permission.

The annotation files released in this repository remain available under the license stated in Section 9.

---

9. Results Summary

OmniVDCBench evaluates 15 representative models across visual-only, audio-only, and joint audio-visual settings.

Metric	Best Proprietary	Best Open-Weight
Overall Accuracy	74.36% (Gemini-3.1 Pro)	55.96% (Qwen3-Omni)

The strongest audio-only baseline, WhisperX, reaches 54.08% Overall Accuracy, indicating that speech transcripts support a substantial subset of the questions. Fine-grained-detail questions are the most challenging across models. A separate expert evidence-writing study on the ICH subset achieves 98.20% Overall Accuracy, providing a human evidence reference. See the paper for the full per-model and per-dimension tables.

---

10. License and Responsible Use

This dataset card uses CC-BY-NC-SA-4.0 as the recommended license for the annotation files.

The dataset is intended for non-commercial research and evaluation. Users should comply with applicable copyright, platform, institutional, and legal requirements when accessing or using video files. The dataset maintainers do not claim ownership over third-party raw videos. If any copyright holder believes that a video should be removed or access restricted, please contact the maintainers through the repository issue or discussion channel.

---

11. Citation

If you use OmniVDCBench in academic work, please cite the associated paper. Before the paper is available, you may cite the dataset as:

@misc{omnivdcbench2026,
  title        = {OmniVDCBench: Caption-as-Evidence Evaluation of Audio-Visual Dense Captioning},
  author       = {Anonymous Authors},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  note         = {1,253 videos, 4,794 caption-anchored evaluation items. Raw video files are released separately after paper acceptance.}
}

---

12. Acknowledgements

We thank all contributors involved in video collection, metadata organization, caption generation and fusion, QA construction, and benchmark packaging. OmniVDCBench is released anonymously by its maintainers to support reproducible academic evaluation of audio-visual dense captioning.