LAE : Long-Tailed Age Estimation

Facial age estimation is an important yet very challenging problem in computer vision. To improve the performance of facial estimation, we first formulate a simple standard baseline and build much strong one by collecting tricks pre-training, data augmentation, model architecture, so on. Compared with baseline, proposed significantly decreases errors. Moreover, long-tailed recognition has been topic datasets, where samples often lack on elderly children. train balanced estimator, propose two-stage training method named Long-tailed Age Estimation (LAE), which decouples learning procedure into representation classification. The effectiveness our approach demonstrated dataset provided organizers Guess Contest 2021.