BanglaFace: A Curated In-the-Wild Bangladeshi Facial Image Dataset for Face Inpainting and Computer Vision

Read full text Summarise with VIRAT
Authors: Istiaq Ahmed Fahad, Md. Saeed Siddik, B M Mainul Hossain
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

DATASET OVERVIEW This dataset consists of 1,625 unique high-quality facial images of Bangladeshi demographics extracted from 45 YouTube videos across 9 content categories. All of these images were collected from publicly available videos which has been enlisted in video_metadat.csv. No additional personal or confidential information is included. These images are curated for face inpainting, facial recognition, and generative modeling. The dataset features both raw in-the-wild and geometrically aligned face representations with comprehensive metadata. SUMMARY STATISTICS OF DATASET - Total Images: 1,625 unique individuals (deduplicated from 80,137 raw detections) - Source Videos: 45 across 9 categories - Total Duration: 22 hours 22 minutes 24 seconds - Demographics: Bangladeshi population RAW VS. ALIGNED IMAGES - Raw Faces: Original crops from video frames via MTCNN detection, retaining natural pose and scale variations for authentic in-the-wild facial data. - Aligned Faces: Standardized 112x112 pixel images with facial landmarks normalized to consistent positions, facilitating model training on facial identity features. PROCESSING PARAMETERS - Face Detector: MTCNN with 0.85 confidence threshold - Minimum Face Resolution: 100x100 pixels - Deduplication: ResNet-34 embeddings (128-dim) with 0.54 similarity threshold - Blur Filtering: Laplacian Variance metric, threshold 80.0 - Frame Sampling: 1 frame per 10 frames - Video Resolution: Minimum 720p CATEGORY DISTRIBUTION - Festival (9 videos): 247 images, 15.20% - University (9): 319 images, 19.63% - Travel (15): 642 images, 39.51% - Market (4): 207 images, 12.74% - Public Events (3): 162 images, 9.97% - Lifestyle (1): 14 images, 0.86% - Education (1): 23 images, 1.42% - Historical Place (1): 8 images, 0.49% - Documentaries (2): 3 images, 0.18% ORGANIZATION Sequential IDs (000001-001625) are organized by category, video, and then by frame order. The image_metadata.csv and video_metadata provide complete traceability, linking each image to the source video, category, and frame ID. The directory structure is shown as follows: BanglaFace_Dataset/ ├── images/ # 1,625 raw face images (000001.jpg - 001625.jpg) ├── aligned_images/ # 1,625 aligned face images (000001.jpg - 001625.jpg) └── metadata/ └── image_metadata.csv # Complete image metadata └── video_metadata.csv # Complete video metadata USE CASES This dataset is suitable for: - Face inpainting and completion tasks - Facial recognition and verification systems - Face alignment and landmark detection - Generative model training (GANs, VAEs) - Face attribute analysis - Demographic studies on facial characteristics

Fields of Study & Subjects

Face recognition and analysis Generative Adversarial Networks and Image Synthesis Face and Expression Recognition

Publication details

Journal / source
Mendeley Data
Publication date
20 March 2026
Language
EN
Document type
Dataset