CVPR2020

Exploring Unlabeled Faces for Novel Attribute Discovery

Hyojin Bahng, Sunghyo Chung, Seungjoo Yoo, Jaegul Choo

Abstract

Input Unpaired & unlabeled translation results (a) Ethnicity / Skin color (b) Hair color (c) Age / Facial hair (d) Additional results Figure 1 : Given raw, unlabeled data, our algorithm discovers novel facial attributes and performs high-quality multi-domain image translation. All results are based on newly-found attributes from our algorithm (e.g., a wide range of ethnicity, skin and hair color, age, facial hair, accessories, and makeup). We did not use any pre-defined attribute labels to generate the results.