This is a face attribute recognition benchmark with mask robustness. And it is divided into two parts, the first part is generated by performing mask synthesis on the existing face attribute recognition datasets CelebA(41 attributes, 199,899 images) and MAAD-Face(48 attributes, 3,093,309 images). The second part is constructed by manually labeling the existing masked face dataset(54 attributes, 11,245 images). In addition to the dataset proposed above, our dataset also includes the original CelebA and MAAD-Face, so our dataset includes faces with and without masks. Based on the degree of mask’s influence on face attribute recognition task, we propose three new evaluation metrics to effectively measure the algorithm’s robustness on masked face attribute recognition task. And they will be published soon...
