Facial recognition and soft biometrics systems often face difficulties in the presence of facial hair, particularly beards, which significantly modify the structure and texture of the face. Existing facial datasets rarely include beard-specific annotations, limiting the robustness of models in real-world conditions. To address this limitation, we introduce BFSET (BeardFaceSet), a curated dataset comprising 4800 fully annotated facial images featuring various beard styles, densities, and poses. Each image includes precise bounding box coordinates of the beard region, enabling targeted analysis and learning. A statistical evaluation based on entropy, GLCM, and HOG features confirms the visual diversity and texture complexity of the dataset. In addition, benchmarking experiments using BFSET demonstrated improved beard detection accuracy (mAP ≈ 0.93) and recognition performance when integrated with existing datasets. BFSET is therefore a valuable resource for the development and evaluation of models that consider beards in facial recognition and soft biometrics.
Elbahri et al. (Tue,) studied this question.