ISL Unconstrained Face Mask Dataset (ISL-UFMD)

This dataset contains ~20k face images. The images are annotated with the following labels: no mask, proper mask, and improper mask. These unconstrained images are collected from various sources on the internet. The ISL-UFMD dataset can be used to develop and evaluate face mask detection and control approaches.


ISL Unconstrained Face Hand Interaction Dataset (ISL-UFHD)

This dataset includes around 30k face images with hand interactions. The images were collected from the internet and annotated with no face-hand interaction and face-hand interaction labels. To the best of our knowledge, this is the only unconstrained face-hand interaction dataset. Researchers can benefit from ISL-UFHD by analyzing face-hand interaction detection techniques and the effects of these types of interactions on COVID-19.


Face Mask Label Dataset (FMLD)

This dataset is an annotation dataset that contains ~63k face image annotations for no mask, correctly and incorrectly worn mask. The annotated images are chosen from the popular unconstrained face datasets MAFA and WIDER Face. In addition to mask annotations, the FMLD also has bounding box coordinates of faces, gender, ethnicity, and pose labels for each face. Researchers can utilize FMLD to develop and improve face mask detection approaches and analyze performance across demographic factors, such as gender, ethnicity of subject.


Twitter Face Mask Detection Dataset (TFMD)

This dataset contains ~16k face images with bounding box annotations. Images also contain mask and no mask class annotations. The entire dataset is collected from Twitter images that are taken in uncontrolled environments. Current face mask datasets mainly were collected from Asian countries where face masks were more commonly used before the COVID-19 pandemic. TFMD contains images of a variety of different ethnicities. Researchers can utilize TFMD to analyze the effect of wearing masks on their face detection models in a realistic scenario.