Hugging Face's logo

Datasets:
26TCA
/
TCA-Bench

Tasks:
Video-Text-to-Text
Modalities:
Text
Video
Formats:
text
Languages:
English
Size:
< 1K
Tags:
video
eccv-2026
audiovisual-captioning
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
377
You are an expert video caption evaluator. Compare the Candidate Caption against the Ground Truth (GT) focusing EXCLUSIVELY on **audio details**.
**Important Constraints:**
- Ignore all visual descriptions.
- Ignore speaker identification errors (who said what).
- Ignore temporal/ordering errors.
Evaluate the Candidate Caption based on the following **2 Dimensions**. Assign a score from **0 to 10** for each dimension.
**Evaluation Dimensions:**
1. **Transcription Accuracy**:
* Accuracy of the verbatim speech content, including preservation of meaning-critical words, fillers, and stutters.
* A meaning-reversing error (e.g., "did" vs. "didn't") or omission of key dialogue warrants a low score.
2. **Tone & Emotion**:
* Accuracy of the described emotional tone, sentiment, and delivery style (e.g., sarcastic, excited, solemn).
* Also covers non-speech audio: music mood, ambient sounds, and sound effects when present in the GT.
**Scoring Criteria (0-10):**
* **0-2 (Critical Failure)**: Meaning is reversed, or the described emotion/tone is opposite to the GT.
* **3-4 (Poor)**: Significant inaccuracies or major omissions that distort meaning or sentiment.
* **5-6 (Fair)**: General gist is correct but with noticeable gaps or minor errors.
* **7-8 (Good)**: Accurate with only minor typos or missing fillers that do not affect meaning.
* **9-10 (Perfect)**: Flawless transcription and precise capture of tone/emotion.
**Input:**
* **Ground Truth:** {{Ground_Truth}}
* **Candidate:** {{Candidate}}
**Output Format:**
Return ONLY a JSON object with the integer scores. No explanations, no extra text.
```json
{
"transcription_accuracy": <int 0-10>,
"tone_and_emotion": <int 0-10>
}
```
You are an expert video caption evaluator. Compare the Candidate Caption against the Ground Truth (GT) focusing EXCLUSIVELY on **audio details**.
**Important Constraints:**
- Ignore all visual descriptions.
- Ignore speaker identification errors (who said what).
- Ignore temporal/ordering errors.
Evaluate the Candidate Caption based on the following **2 Dimensions**. Assign a score from **0 to 10** for each dimension.
**Evaluation Dimensions:**
1. **Transcription Accuracy**:
* Accuracy of the verbatim speech content, including preservation of meaning-critical words, fillers, and stutters.
* A meaning-reversing error (e.g., "did" vs. "didn't") or omission of key dialogue warrants a low score.
2. **Tone & Emotion**:
* Accuracy of the described emotional tone, sentiment, and delivery style (e.g., sarcastic, excited, solemn).
* Also covers non-speech audio: music mood, ambient sounds, and sound effects when present in the GT.
**Scoring Criteria (0-10):**
* **0-2 (Critical Failure)**: Meaning is reversed, or the described emotion/tone is opposite to the GT.
* **3-4 (Poor)**: Significant inaccuracies or major omissions that distort meaning or sentiment.
* **5-6 (Fair)**: General gist is correct but with noticeable gaps or minor errors.
* **7-8 (Good)**: Accurate with only minor typos or missing fillers that do not affect meaning.
* **9-10 (Perfect)**: Flawless transcription and precise capture of tone/emotion.
**Input:**
* **Ground Truth:** {{Ground_Truth}}
* **Candidate:** {{Candidate}}
**Output Format:**
Return a JSON object with the scores and a brief reasoning for each:
```json
{
"transcription_accuracy": {
"score": <int 0-10>,
"reason": "<brief explanation>"
},
"tone_and_emotion": {
"score": <int 0-10>,
"reason": "<brief explanation>"
}
}
```
You are an expert video caption evaluator. Compare the Candidate Caption against the Ground Truth (GT) focusing EXCLUSIVELY on **visual details**.
**CRITICAL INSTRUCTION: IGNORE ALL SPEECH, DIALOGUE, AND AUDIO.**
Do not penalize the candidate if it omits spoken content, attributes speech to the wrong character, or fails to transcribe dialogue. This evaluation is strictly for visual fidelity.
Evaluate the Candidate Caption based on the following **3 Dimensions** derived from the VC4VG framework. Assign a score from **0 to 10** for each dimension.
**Evaluation Dimensions:**
1. **Subject & Action**:
* Accuracy of the primary subjects' descriptions, including quantity, appearance, clothing, accessories, and spatial relationships (positions and interactions).
* Accuracy of movement and temporal changes, including sequential actions (limb movements), movement paths (direction and position changes), and gradual state transitions.
2. **Scene & Atmosphere**:
* Accuracy of the setting and background description, including spatiotemporal attributes (lighting, weather, time-of-day) and geospatial layout (object placement, background scenery).
* Accuracy of the visual style and mood, including emotional ambiance (color grading, lighting mood) and artistic/rendering qualities (e.g., cinematic, grainy, vivid).
3. **Cinematography**:
End of preview. Expand in Data Studio

TCA-Bench

Official ECCV 2026 benchmark release for the paper "Temporal and Cross-modal Alignment for Enhanced Audiovisual Video Captioning."

TCA-Bench is a diagnostic benchmark for audiovisual video captioning. It evaluates base audio/visual perception, audio-visual binding, and cross-modal temporal reasoning using structured ground truth annotations.

The benchmark contains 459 anonymized short videos. All annotation files use the anonymized mp4 filename as id, matching files in videos/.

Files

  • videos.tar.gz: compressed archive containing anonymized video files named tca_bench_000001.mp4 through tca_bench_000459.mp4.
  • gt/captions.json: Stage-1 ground-truth captions.
  • gt/stage-2.json: Stage-2 audio-visual binding lists.
  • gt/stage-3.json: Stage-3 temporal relation lists.
  • scripts/evaluate.py: evaluation script.
  • scripts/prompts/: captioning and judge prompts.
  • requirements.txt: evaluator dependencies.
  • source_manifest.csv: source URL and temporal segment for each anonymized video.

Evaluation

Stage 1 evaluates base perception against gt/captions.json:

  • 1v: visual quality.
  • 1a: audio quality.

Stage 2 evaluates audio-visual binding against gt/stage-2.json.

Stage 3 evaluates temporal reasoning against gt/stage-3.json and reports Temporal F1 with precision and coverage.

scripts/prompts/caption_prompt.txt is the recommended prompt for generating candidate captions.

Extract videos.tar.gz in the repository root before reading video files:

tar -xzf videos.tar.gz

Candidate captions should be a JSON array or JSONL file:

{"id": "tca_bench_000001.mp4", "caption": "Candidate caption text..."}

Run evaluation:

python -m pip install -r requirements.txt
export TCA_EVAL_API_KEY="..."
export TCA_EVAL_BASE_URL="https://openrouter.ai/api/v1"
export TCA_EVAL_MODEL="gpt-4.1"
python scripts/evaluate.py --input path/to/captions.json --name my-model --stage 1v 1a 2 3

Outputs are written to results/<name>/<timestamp>/.

License

This release is for non-commercial research use under CC BY-NC-SA 4.0.

Citation 

@misc{tca_bench_2026,
  title = {Temporal and Cross-Modal Alignment for Enhanced Audiovisual Video Captioning},
  year = {2026}
}
Downloads last month
80
Number of rows:
326
Total file size:
1.07 GB

Collection including 26TCA/TCA-Bench