Collection of 2429 constrained headshots of 277 volunteers for deep learning

• DOI: 10.1101/2020.10.14.337220
• Publication Date: 2020-10-16T05:05:12Z
• Authors: Saki Aoto; Mayumi Hangai; Hitomi Ueno-Yokohata; Aki Ueda; Maki Igarashi; Yoshikazu Ito; Motoko Tsukamoto; Tomoko Jinno; Mika Sakamoto; Yuka Okazaki; Fuyuki Hasegawa; Hiroko Ogata-Kawata; Saki Namura; Kazuaki Kojima; Masao Kikuya; Keiko Matsubara; Kosuke Taniguchi; Kohji Okamura
• Keywords: Feature (linguistics)

Abstract Deep learning has rapidly been filtrating many aspects of human lives. In particular, image recognition by convolutional neural networks inspired numerous studies in this area. Hardware and software technologies as well large quantities data have contributed to the drastic development field. However, application deep is often hindered need for big laborious manual annotation thereof. To experience using compiled us, we collected 2429 constrained headshot images 277 volunteers. The collection face photographs challenging terms protecting personal information; established an online procedure which both informed consent could be obtained. We did not collect information, but issued agreement numbers deal with withdrawal requests. Gender smile labels were manually subjectively annotated only from appearances, final determined majority among our team members. Rotated, trimmed, resolution-reduced, decolorized, matrix-formed allowed publicly released. Moreover, simplified feature vectors sciences performed gender building based on Inception V3 model pre-trained ImageNet demonstrate usefulness dataset.