# CameraBench Binary Evaluation Dataset

A balanced VQA dataset for evaluating camera motion understanding in videos.

## 📊 Dataset Statistics

- **Total Questions**: 232
- **Unique Videos**: 115
- **Unique Questions**: 15
- **Yes Answers**: 116 (50.0%)
- **No Answers**: 116 (50.0%)
- **Balance Ratio**: 1.00
- **Total Size**: 121.22 MB (0.12 GB)
- **Average Video Size**: 1.05 MB

## 🎯 Task Categories

This dataset covers various camera motion tasks including:
- **Static**: 39 questions
- **Move In**: 30 questions
- **Pan Right**: 23 questions
- **Pan Left**: 23 questions
- **Move Out**: 20 questions
- **Roll Counterclockwise**: 20 questions
- **Roll Clockwise**: 19 questions
- **Move Down**: 17 questions
- **Tilt Up**: 16 questions
- **Move Up**: 16 questions
- **Move Right**: 14 questions
- **Move Left**: 14 questions
- **Tilt Down**: 13 questions
- **Zoom In**: 13 questions
- **Zoom Out**: 12 questions


## 📝 Dataset Format

The dataset consists of:
- `videos/`: Directory containing all MP4 video files
- `metadata.jsonl`: JSONL file with question annotations

Each record in `metadata.jsonl` contains:
- `video_name`: Original video filename
- `video_path`: Relative path to video file (e.g., `videos/video.mp4`)
- `question`: Binary question about camera motion
- `label`: Answer ("Yes" or "No")
- `task`: Task category
- `label_name`: Detailed label identifier

## 🎬 Sample Questions and Videos

Below are animated GIF previews of sample videos from the dataset:



## 🚀 Usage

### Loading the Dataset

```python
import json
import os

# Load metadata
metadata = []
with open("metadata.jsonl", "r") as f:
    for line in f:
        metadata.append(json.loads(line))

# Access a sample
sample = metadata[0]
print(f"Question: {sample['question']}")
print(f"Answer: {sample['label']}")
print(f"Task: {sample['task']}")
print(f"Video path: {sample['video_path']}")
```

### Downloading the Dataset

Download the entire dataset using huggingface-cli or git:

```bash
# Using huggingface-cli
huggingface-cli download tuhink/cambench_binary_eval --repo-type dataset --local-dir ./cambench_data

# Or using git
git clone https://huggingface.co/datasets/tuhink/cambench_binary_eval
```

This will download all videos and metadata to your local machine.

### Loading Videos

```python
import json
import cv2

# Load metadata
with open("metadata.jsonl", "r") as f:
    metadata = [json.loads(line) for line in f]

# Load a video
sample = metadata[0]
video_path = sample['video_path']  # e.g., "videos/video_name.mp4"

# Use OpenCV to read the video
cap = cv2.VideoCapture(video_path)
while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break
    # Process frame
    pass
cap.release()
```

### Batch Processing

For evaluation tasks:

```python
import json

# Load all questions
with open("metadata.jsonl", "r") as f:
    dataset = [json.loads(line) for line in f]

correct = 0
total = 0

for sample in dataset:
    video_path = sample['video_path']
    question = sample['question']
    ground_truth = sample['label']
    
    # Your model inference here
    # prediction = your_model(video_path, question)
    
    # if prediction == ground_truth:
    #     correct += 1
    # total += 1

# accuracy = correct / total if total > 0 else 0
# print(f"Accuracy: {accuracy:.2%}")
```

### Using with HuggingFace Datasets Library

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("tuhink/cambench_binary_eval")

# Access samples
for sample in dataset['train']:
    print(f"Question: {sample['question']}")
    print(f"Answer: {sample['label']}")
    print(f"Video: {sample['video_path']}")
```

## 📊 Evaluation

This dataset is designed for binary classification tasks. Evaluate your model using:
- Accuracy
- Precision/Recall
- F1 Score
- Per-task performance

## 📄 License

Please refer to the original CameraBench dataset for licensing information.

## 🙏 Citation

If you use this dataset, please cite the original CameraBench paper.

## 📧 Contact

For questions or issues, please open an issue on the repository.

---

**Note**: All videos are provided in original MP4 format. The dataset maintains temporal dynamics for accurate camera motion evaluation.