OF10K: Digging Into a New Dataset Containing Diversified Naturally-Occluded Facial Images

Facial recognition is increasingly applied across diverse real-world contexts; however, recognizing partially visible faces remains underexplored due to the absence of comprehensive datasets reflecting natural obstructions. To address this gap, we present OF10K (Occluded Facial 10K), a large and diverse dataset containing 9,540 facial images with various natural occlusions, such as caps, sunglasses, masks, and combined mask–sunglass coverings. Images were collected from both direct camera captures and publicly available sources to ensure diversity in lighting, pose, age group, and environment. Using OF10K, we evaluate the performance of state-of-the-art deep learning architectures, including ResNet, VGG, and YOLO, to benchmark recognition capability under partial visibility. Our analyses reveal both the strengths and limitations of these models in real-world occlusion scenarios. To the best of our knowledge, this is the first study to introduce a large-scale dataset of naturally obstructed facial images along with a systematic evaluation of popular deep learning models. OF10K establishes a new benchmark for advancing research in occlusion-aware facial recognition.