Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
µö·¯´× ±â¹Ý ÇÇ»çü ½Ã¼± ÃßÀûÀ» ÅëÇÑ ÀÚµ¿ ÁÖ¼® »ý¼º ½Ã½ºÅÛ |
¿µ¹®Á¦¸ñ(English Title) |
|
ÀúÀÚ(Author) |
Á¤ÁöÀº
ÃÖ¿ë¼®
Ji Eun Jeong
Yong Suk Choi
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 27 NO. 03 PP. 0157 ~ 0162 (2021. 03) |
Çѱ۳»¿ë (Korean Abstract) |
ÇÇ»çüÀÇ ½Ã¼± ÃßÀû(Gaze Following)Àº ´ÜÀÏ À̹ÌÁö¿¡¼ ÇÇ»çüÀÇ ½Ã¼±ÀÌ ÀÀ½ÃÇÏ´Â ÁöÁ¡À» ŽÁöÇÑ´Ù. µö·¯´× ±â¹ÝÀÇ ±âÁ¸ ¿¬±¸´Â ´Ü¼øÈ÷ ½Ã¼±ÀÇ °¢µµ¸¦ ÃßÁ¤Çϰųª, ½º¸¶Æ®Æù°ú °°Àº ±â±â ½ºÅ©¸° ³»ºÎ ÀÇ ÀÀ½ÃÁ¡À» ÃßÁ¤ÇϹǷΠ¾î¶² °´Ã¼¸¦ º¸´ÂÁö¿¡ ´ëÇÑ Á¤º¸¸¦ ¾òÀ» ¼ö ¾ø´Ù´Â ÇÑ°è°¡ ÀÖ´Ù. º» ³í¹®¿¡¼´Â ÃÖÃÊ·Î µö·¯´× ¸ðµ¨À» È°¿ëÇÏ¿© ÇÇ»çüÀÇ ½Ã¼±À» ÃßÀûÇÏ°í ¡®J°¡ ³ØŸÀ̸¦ º»´Ù.¡¯¿Í °°ÀÌ ÀÚµ¿À¸·Î ÁÖ¼®À» »ý¼ºÇÏ´Â ½Ã½ºÅÛÀ» Á¦¾ÈÇÑ´Ù. ½Ã½ºÅÛÀº Àüó¸® ¸ðµâ, ½Ã¼± ÃßÀû ¸ðµâ, ÈÄó¸® ¸ðµâ·Î ±¸¼ºµÇ¸ç, Àüó¸® ¸ðµâ ¿¡¼ Àι°À» ÀνÄÇÏ°í µö·¯´× ¸ðµ¨ÀÇ ÀÔ·ÂÀ» »ý¼ºÇÑ´Ù. ½Ã¼± ÃßÀû ¸ðµâ¿¡¼´Â µö·¯´× ¸ðµ¨ÀÌ ÇÇ»çüÀÇ ÀÀ½Ã ÁöÁ¡ÀÌ Ç¥½ÃµÈ È÷Æ®¸Ê(heatmap)À» »ý¼ºÇÑ´Ù. ÈÄó¸® ¸ðµâ¿¡¼´Â ¿ì¸®°¡ Á¦¾ÈÇÏ´Â °´Ã¼ ¼±Åà ¾Ë°í¸®Áò¿¡ ÀÇÇØ ÀÀ½Ã ÁöÁ¡¿¡ ÀÖ´Â °´Ã¼¸¦ ÀνÄÇÏ°í ÁÖ¼®À» »ý¼ºÇÑ´Ù. Á¦¾ÈµÈ ½Ã½ºÅÛÀº ¸®Å×Àϸµ ¹× Çмú ¸ñÀûÀÇ ´ë±Ô¸ð ¸ÞŸµ¥ÀÌÅ͸¦ È¿À²ÀûÀ¸·Î »ý¼ºÇÏ´Â µ¥ È°¿ëµÉ ¼ö ÀÖ´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
Gaze following is the task of detecting the point of attention in a single image at which a subject's gaze is staring. The existing deep learning methods have a limitation in that they cannot determine which specific object is the target of the gaze angle, because these methods simply estimate the gaze angle, or estimate the gaze point within the screen of a device, such as a smart phone. In this paper, we propose a novel system that infers where a subject is looking by using a deep learning model and automatically generating annotations such as 'J is looking at the tie.' The proposed system consists of a pre-processing module, a gaze following module, and a post-processing module. In the pre-processing module, our system recognizes faces and generates inputs for the deep learning model. In the gaze following module, the deep neural network generates a heatmap for the point at which the subject is looking. In the post-processing module, the proposed object selection algorithm determines the object of the gaze point, and annotations are then generated. The proposed system can be used to generate large-scale metadata for retailing and academic purposes.
|
Å°¿öµå(Keyword) |
µö·¯´×
½Ã¼± ÃßÀû
½Ã¼± ÃßÁ¤
ÁÖ¼® »ý¼º
deep learning
gaze following
gaze estimation
annotation generation
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|