| # CameraBench Binary Evaluation Dataset |
|
|
| A balanced VQA dataset for evaluating camera motion understanding in videos. |
|
|
| ## π 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 |
|
|
| ## π― Task Categories |
|
|
| This dataset covers various camera motion tasks including: |
| - **Static**: 39 questions |
| - **Move In**: 27 questions |
| - **Pan Right**: 26 questions |
| - **Roll Counterclockwise**: 22 questions |
| - **Pan Left**: 21 questions |
| - **Move Left**: 20 questions |
| - **Roll Clockwise**: 18 questions |
| - **Move Up**: 18 questions |
| - **Move Out**: 18 questions |
| - **Tilt Up**: 17 questions |
| - **Move Right**: 14 questions |
| - **Move Down**: 14 questions |
| - **Zoom In**: 14 questions |
| - **Tilt Down**: 13 questions |
| - **Zoom Out**: 12 questions |
|
|
|
|
| ## π Dataset Format |
|
|
| Each record contains: |
| - `video_name`: Original video filename |
| - `video`: Video file (MP4 format, original quality) |
| - `question`: Binary question about camera motion |
| - `label`: Answer ("Yes" or "No") |
| - `task`: Task category |
| - `label_name`: Detailed label identifier |
|
|
| ## π¬ Sample Questions and Videos |
|
|
| Below are animated GIF previews of sample videos from the dataset: |
|
|
|
|
|
|
| ## π Usage |
|
|
| ### Loading the Dataset |
|
|
| ```python |
| from datasets import load_dataset |
| |
| # Load the dataset |
| dataset = load_dataset("cambench_binary_eval") |
| |
| # Access a sample |
| sample = dataset['train'][0] |
| print(f"Question: {sample['question']}") |
| print(f"Answer: {sample['label']}") |
| print(f"Task: {sample['task']}") |
| |
| # The video field contains the path to download |
| video_file = sample['video'] |
| ``` |
|
|
| ### Downloading Videos |
|
|
| Videos are embedded in the dataset. To download and use them: |
|
|
| ```python |
| from datasets import load_dataset |
| import os |
| import shutil |
| |
| # Load the dataset |
| dataset = load_dataset("cambench_binary_eval") |
| |
| # Download all videos to a local directory |
| output_dir = "downloaded_videos" |
| os.makedirs(output_dir, exist_ok=True) |
| |
| for idx, sample in enumerate(dataset['train']): |
| video_name = sample['video_name'] |
| video_data = sample['video'] # This contains the video file data |
| |
| # Save video to local disk |
| local_path = os.path.join(output_dir, video_name) |
| |
| # If video_data is a file path (during local testing) |
| if isinstance(video_data, str) and os.path.exists(video_data): |
| shutil.copy2(video_data, local_path) |
| else: |
| # Video data from HuggingFace - write bytes to file |
| with open(local_path, 'wb') as f: |
| f.write(video_data) |
| |
| print(f"Downloaded: {video_name} -> {local_path}") |
| |
| print(f"All {len(dataset['train'])} videos downloaded to {output_dir}/") |
| ``` |
|
|
| ### Accessing Individual Videos |
|
|
| To download a specific video by name: |
|
|
| ```python |
| from datasets import load_dataset |
| import os |
| |
| dataset = load_dataset("cambench_binary_eval") |
| |
| # Find and download a specific video |
| target_video = "your_video_name.mp4" |
| for sample in dataset['train']: |
| if sample['video_name'] == target_video: |
| video_data = sample['video'] |
| |
| # Save to current directory |
| with open(target_video, 'wb') as f: |
| f.write(video_data) |
| |
| print(f"Downloaded: {target_video}") |
| break |
| else: |
| print(f"Video {target_video} not found in dataset") |
| ``` |
|
|
| ### Batch Processing |
|
|
| For evaluation tasks: |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("cambench_binary_eval") |
| |
| correct = 0 |
| total = 0 |
| |
| for sample in dataset['train']: |
| video_path = sample['video'] |
| 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%}") |
| ``` |
|
|
| ## π₯ Alternative: Download Full Dataset |
|
|
| To download the entire dataset with all videos at once: |
|
|
| ```bash |
| # Using huggingface-cli |
| huggingface-cli download cambench_binary_eval --repo-type dataset --local-dir ./cambench_data |
| |
| # Or using Python |
| from huggingface_hub import snapshot_download |
| snapshot_download(repo_id="cambench_binary_eval", repo_type="dataset", local_dir="./cambench_data") |
| ``` |
|
|
| This will download all videos and data files to your local machine. |
|
|
| ## π Evaluation |
|
|
| This dataset is designed for binary classification tasks. Evaluate your model using: |
| - Accuracy |
| - Precision/Recall |
| - F1 Score |
| - Per-task performance |
|
|
| ## π License |
|
|
| Please refer to the original CameraBench dataset for licensing information. |
|
|
| ## π Citation |
|
|
| If you use this dataset, please cite the original CameraBench paper. |
|
|
| ## π§ Contact |
|
|
| For questions or issues, please open an issue on the repository. |
|
|
| --- |
|
|
| **Note**: Videos are provided in MP4 format. Large videos (>50MB) are automatically compressed to ensure smooth downloading and processing. All videos maintain their original temporal dynamics for accurate camera motion evaluation. |
|
|