README.md

---
language:
- en
task_categories:
- video-text-to-text
---

<!-- # TimeBlind Benchmark -->

<!-- TimeBlind: A video VQA benchmark for evaluating temporal understanding in vision-language models -->
<div align="center">
  <strong style="font-size: 28px; line-height: 1.2; font-family: Arial, sans-serif;">
    TimeBlind: A Spatio-Temporal Compositionality Benchmark for Video LLMs
  </strong>
</div>

<div align="center"> Baiqi Li<sup>1</sup>, Kangyi Zhao<sup>2</sup>, Ce Zhang<sup>1</sup>, Chancharik Mitra<sup>3</sup>, Jean de Dieu Nyandwi<sup>3</sup>, Gedas Bertasius<sup>1</sup> </div> <div align="center"> <sup>1</sup>University of North Carolina at Chapel Hill&nbsp;&nbsp; <sup>2</sup>University of Pittsburgh&nbsp;&nbsp; <sup>3</sup>Carnegie Mellon University </div>


<div align="center">

[🏠**Home Page**](https://baiqi-li.github.io/timeblind_project/) | [🤗**HuggingFace**](https://huggingface.co/datasets/BaiqiL/TimeBlind) | [**📖Paper**](https://huggingface.co/papers/2602.00288) | [🖥️ **Code**](https://github.com/Baiqi-Li/TimeBlind)

</div>

## Introduction
Fine-grained spatio-temporal understanding is essential for video reasoning and embodied AI. TimeBlind is a diagnostic benchmark for compositional spatio-temporal understanding. Inspired by cognitive science, TimeBlind categorizes fine-grained temporal understanding into three levels: recognizing atomic events, characterizing event properties, and reasoning about event interdependencies. It leverages a minimal-pairs paradigm where video pairs share identical static visual content but differ solely in temporal structure, utilizing complementary questions to neutralize language priors.

## Setup

```bash
git clone https://github.com/Baiqi-Li/TimeBlind.git
cd TimeBlind
git clone https://huggingface.co/datasets/BaiqiL/TimeBlind
```

## Data Format

Each sample in `TimeBlind/data.jsonl` contains:
- `index`: unique sample index (0, 1, 2, ...)
- `video_path`: path to video file (e.g., `TimeBlind/videos/vid_00000_0.mp4`)
- `question`: the question
- `answer`: the ground truth answer
- `type`: `"yes_no"` or `"multiple_choice"`

## Evaluation

see evaluate.py in our github page for more details!

```python
import json
from utils import _load_json_list, build_answers, get_scores, add_question_suffix

data = _load_json_list("TimeBlind/data.jsonl")
predictions = []

for sample in data:
    video_path = sample["video_path"]  
    question = add_question_suffix(sample["question"], sample["type"])
    
    # Replace with your model inference
    model_output = your_model(video_path, question)
    
    predictions.append({
        "index": sample["index"],
        "video_path": video_path,
        "question": question,
        "model_output": model_output,
    })

json.dump(predictions, open("predictions.json", "w"), indent=2)

answers = build_answers(predictions, data)
scores = get_scores(answers)
print(scores)  # {'Q_Acc': ..., 'V_Acc': ..., 'Acc': ..., 'I_Acc': ...}
```

## Metrics
I-Acc serves as our primary metric.

- **Acc**: Binary VQA accuracy
- **Q_Acc**: Question accuracy
- **V_Acc**: Video accuracy
- **I_Acc**: Instance accuracy (the primary metric in our paper)

# Copyright & Infringement Notice
The data provided in this benchmark is intended for academic research purposes only. We respect the intellectual property rights of the content creators.

If you believe that any content in this dataset infringes upon your rights, please contact us at [baiqili@unc.cs.edu] and we will remove the relevant content immediately.