Picking out the bad apples: unsupervised biometric data filtering for refined age estimation

Introduction of large training datasets was essential for the recent advancement and success deep learning methods. Due to difficulties related biometric data collection, facial image with trait labels are scarce usually limited in terms size sample diversity. Web-scraping approaches automatic collection can produce amounts weakly labeled noisy data. This work is focused on picking out bad apples from web-scraped by automatically removing erroneous samples that impair their usability. The unsupervised filtering method presented this greatly reduces label noise levels Experiments two state-of-the-art demonstrate effectiveness proposed respect real apparent age estimation based five different Furthermore, we apply method, together a newly devised strategy merging multiple datasets, collected three major web-based sources (i.e., IMDb, Wikipedia, Google) derive new Biometrically Filtered Famous Figure Dataset or B3FD. dataset, which made publicly available, enables considerable performance gains all tested methods tasks. highlights importance quality compared quantity selection method.