nikhilny25/ff-images-dataset

--- license: cc-by-nc-4.0 task_categories: - image-classification tags: - deepfake-detection - faceforensics - computer-vision - binary-classification size_categories: - 100K<n<1M --- # FaceForensics++ Image Dataset This dataset contains preprocessed images from the FaceForensics++ benchmark for deepfake detection. ## Dataset Description - **Total Images:** 223,919 - **Real Images:** 32,000 - **Fake Images:** 191,919 - **Imbalance Ratio:** 6.00:1 (fake:real) ### Categories | Category | Count | |----------|-------| | original | 32,000 | | Deepfakes | 32,000 | | Face2Face | 32,000 | | FaceSwap | 32,000 | | NeuralTextures | 32,000 | | FaceShifter | 32,000 | | DeepFakeDetection | 31,919 | ## Usage ```python from datasets import load_dataset # Load the dataset dataset = load_dataset("RohanRamesh/ff-images-dataset") # Access splits train_data = dataset['train'] val_data = dataset['validation'] test_data = dataset['test'] # Example: iterate over training data for sample in train_data: image = sample['image'] # PIL Image label = sample['label'] # 0 = FAKE, 1 = REAL category = sample['category'] # e.g., 'original', 'Deepfakes', etc. ``` ## Dataset Structure Each sample contains: - `image`: The face image (PIL Image) - `label`: Binary label (0 = FAKE, 1 = REAL) - `category`: Original category (original, Deepfakes, Face2Face, FaceSwap, FaceShifter, NeuralTextures, DeepFakeDetection) - `video_id`: Source video identifier - `frame_number`: Frame number within the video - `label_text`: Text label ("REAL" or "FAKE") ## Splits The dataset is split by video ID to prevent data leakage: - **Train:** 80% of videos - **Validation:** 10% of videos - **Test:** 10% of videos ## Citation If you use this dataset, please cite the original FaceForensics++ paper: ```bibtex @inproceedings{roessler2019faceforensicspp, author = {Rossler, Andreas and Cozzolino, Davide and Verdoliva, Luisa and Riess, Christian and Thies, Justus and Niessner, Matthias}, title = {FaceForensics++: Learning to Detect Manipulated Facial Images}, booktitle = {International Conference on Computer Vision (ICCV)}, year = {2019} } ```--- license: 知识共享署名-非商业性使用4.0（CC BY-NC 4.0） task_categories: - 图像分类（image-classification） tags: - 深度伪造检测（deepfake-detection） - FaceForensics++（面部伪造基准） - 计算机视觉（computer-vision） - 二分类（binary-classification） size_categories: - 100,000 < 样本量 < 1,000,000 --- # FaceForensics++ 图像数据集 本数据集包含面向深度伪造检测的FaceForensics++基准测试的预处理人脸图像。 ## 数据集概况 - **总图像数：** 223,919 - **真实图像数：** 32,000 - **伪造图像数：** 191,919 - **样本不平衡比例：** 6.00:1（伪造样本:真实样本） ### 类别分布 | 类别名称 | 样本数量 | |----------|----------| | 原始图像（original） | 32,000 | | Deepfakes | 32,000 | | Face2Face | 32,000 | | FaceSwap | 32,000 | | NeuralTextures | 32,000 | | FaceShifter | 32,000 | | DeepFakeDetection | 31,919 | ## 使用方法 python from datasets import load_dataset # 加载数据集 dataset = load_dataset("RohanRamesh/ff-images-dataset") # 访问数据集划分 train_data = dataset['train'] val_data = dataset['validation'] test_data = dataset['test'] # 示例：遍历训练集数据 for sample in train_data: image = sample['image'] # PIL图像（PIL Image） label = sample['label'] # 0 代表伪造样本（FAKE），1 代表真实样本（REAL） category = sample['category'] # 例如：'original'、'Deepfakes' 等 ## 数据集结构 每个样本包含以下字段： - `image`：人脸图像（PIL图像（PIL Image）） - `label`：二分类标签（0 代表伪造样本，1 代表真实样本） - `category`：原始类别（可选值为：original、Deepfakes、Face2Face、FaceSwap、FaceShifter、NeuralTextures、DeepFakeDetection） - `video_id`：来源视频标识符 - `frame_number`：视频内的帧序号 - `label_text`：文本标签（"REAL" 或 "FAKE"） ## 数据集划分 本数据集按照视频ID进行划分，以避免数据泄露： - **训练集：** 80%的视频样本 - **验证集：** 10%的视频样本 - **测试集：** 10%的视频样本 ## 引用说明 若您使用本数据集，请引用原始FaceForensics++学术论文： bibtex @inproceedings{roessler2019faceforensicspp, author = {Rossler, Andreas and Cozzolino, Davide and Verdoliva, Luisa and Riess, Christian and Thies, Justus and Niessner, Matthias}, title = {FaceForensics++: Learning to Detect Manipulated Facial Images}, booktitle = {国际计算机视觉大会（International Conference on Computer Vision, ICCV)}, year = {2019} }