BanglaFace: An In-the-Wild Bangladeshi Facial Image Dataset

BanglaFace is a curated facial image dataset containing 2,500 high-quality face images of Bangladeshi individuals collected from three publicly available image-sharing platforms: Flickr, Pexels, and Unsplash. The images were retrieved using 13 carefully designed search queries covering a broad range of demographic groups, occupations, and everyday scenes found in Bangladesh. All source images originate from permissively licensed collections, including Creative Commons licenses, the Pexels License, and the Unsplash License. The dataset is intended for research in face recognition, face alignment, face restoration, face inpainting, facial analysis, and generative modeling. Both original in-the-wild face crops and geometrically aligned face images are provided together with comprehensive metadata. Summary Statistics Metric Value Total Images 2,500 (selected from ~7,800 downloaded images and ~3253 unique images after deduplication) Source Platforms 3 (Flickr, Pexels, Unsplash) Search Categories 13 Target Population Bangladeshi individuals Gender Distribution Male: 1,748 (69.9%), Female: 752 (30.1%) Raw and Aligned Face Images Raw Face Images (160×160): Faces are detected using YOLOv8n-face and cropped with a context-preserving bounding box expansion factor of 1.8×. These images retain natural pose, scale, illumination, and background variations, making them suitable for in-the-wild facial analysis. Aligned Face Images (112×112): Facial landmarks are extracted using the MediaPipe FaceLandmarker, and faces are normalized to the standard InsightFace/ArcFace template through a five-point similarity transformation. The aligned images provide a consistent facial geometry for recognition and representation learning. Processing Pipeline Parameter Value Face Detection Model YOLOv8n-face (Ultralytics) Detection Confidence Threshold 0.75 Minimum Accepted Face Size 50 × 50 pixels Bounding Box Expansion 1.8× Acceptable Aspect Ratio 0.6–1.2 Image Quality Filtering Brightness: 40–220; Contrast: σ ≥ 25 Raw Face Resolution 160 × 160 pixels Image Interpolation INTER_AREA JPEG Quality 95 Attribute Classification FairFace (ResNet-34, ONNX) Landmark Detection MediaPipe FaceLandmarker (478 landmarks) Face Alignment Five-point similarity transform using OpenCV estimateAffinePartial2D Alignment Success Rate 97.5% (2,437 of 2,500 images) Duplicate Removal MD5 hash-based exact image deduplication Manual Verification Interactive web-based annotation tool with manual gender correction Category Distribution Category Count Percentage Students 516 20.64% Smiling Portrait 288 11.52% Village Portrait 224 8.96% Market People 219 8.76% Garment Worker 216 8.64% People Portrait 210 8.40% Headshot 186 7.44% Urban People 156 6.24% Urban Portrait 131 5.24% Farmer 120 4.80% Elderly 96 3.84% Rickshaw Puller 76 3.04% Fisherman 62 2.48% Total 2,500 100.00% Age and Gender Distribution Age Range Male Female Total Percentage 0–2 5 1 6 0.2% 3–9 388 137 525 21.0% 10–19 332 120 452 18.1% 20–29 247 268 515 20.6% 30–39 207 63 270 10.8% 40–49 194 63 257 10.3% 50–59 48 9 57 2.3% 60–69 224 39 263 10.5% 70+ 103 52 155 6.2% Total 1,748 752 2,500 100.0% Dataset Organization Each image is assigned a unique sequential identifier ranging from img_000001.jpg to img_002500.jpg. The dataset includes raw face crops, aligned face images, metadata, category mappings, and licensing information. bd-face-core/ ├── raw/ # Raw face crops (160×160) │ ├── img_000001.jpg │ ├── img_000002.jpg │ └── ... ├── aligned/ # Aligned face images (112×112) │ ├── img_000001.jpg │ ├── img_000002.jpg │ └── ... ├── metadata.csv # Image-level metadata ├── category_mapping.json # Search scene to category mapping └── license.txt # Source license informationbd-face-raw-images.zip/ # Raw downloaded images Metadata Fields Field Description image_id Unique numeric identifier (1–2500) filename Standardized filename (e.g., img_000001.jpg) source_platform Image source platform (Flickr, Pexels, or Unsplash) search_scene Original search query used during image collection category Assigned semantic category original_filename Original filename from the downloaded source original_width Width of the raw face image (160 pixels) original_height Height of the raw face image (160 pixels) aligned_width Width of the aligned face image (112 pixels) aligned_height Height of the aligned face image (112 pixels) license License associated with the original source image Applications The BanglaFace dataset can be used for a wide range of computer vision and machine learning tasks, including: Face recognition and face verification Face alignment and landmark localization Face restoration and face inpainting Generative modeling using GANs, VAEs, and diffusion models Facial attribute prediction Demographic and fairness analysis Bias-aware evaluation and development of facial recognition systems