Can one create a synthetic dataset by also using images from CASIA dataset?
Posted by: HectorPenades @ Feb. 23, 2024, 1:17 p.m.Yes, for this challenge you can use synthetic images created with a model trained with the CASIA dataset.
Posted by: FRCSyn @ Feb. 26, 2024, 9:39 a.m.Regarding this point, and as indicated in the instructions of the challenge, we would like to remark on the rules that "Synthetic face images can be generated using any state-of-the-art methods, no restrictions." These state-of-the-art synthesis models can be trained with any dataset, including the CASIA-WebFace dataset. However, it is important to remark two points: 1) the evaluation datasets considered in the challenge (i.e., BUPT, AgeDB, CFP-FP and ROF) cannot be used for anything. They must be considered only for the final evaluation of the face recognition model; 2) real data cannot be directly used to generate synthetic data (e.g., via data augmentation techniques), but rather must be employed to train a generative framework that subsequently synthesizes data.
Posted by: FRCSyn @ Feb. 26, 2024, 2:53 p.m.Can we get some information (not images, but something like "style bank" as in DCFace paper) from the real dataset and use it to generate synthetic dataset?
If yes, what information could we use, and from which dataset ? (only CASIA WebFace, or any dataset?)