Electronics and Telecommunications Trends 인공지능을이용한 3D 콘텐츠기술동향및향후전망 Recent Trends and Prospects of 3D Content Using Artificial Intelligence Technology 이승욱 (S.W. Lee, tajinet@etri.re.kr) 황본우 (B.W. Hwang, bwhwang@etri.re.kr) 임성재 (S.J. Lim, sjlimg@etri.re.kr) 윤승욱 (S.U. Yoon, suyoon@etri.re.kr) 김태준 (T.J. Kim, taejoonkim@etri.re.kr) 김기남 (K.N. Kim, rlskal@etri.re.kr) 김대희 (D.H Kim, kdh60243@etri.re.kr) 박창준 (C.J. Park, chjpark@etri.re.kr) CG/Vision연구실책임연구원 CG/Vision연구실책임연구원 CG/Vision연구실책임연구원 CG/Vision연구실선임연구원 CG/Vision연구실선임연구원 CG/Vision연구실선임기술원 CG/Vision연구실위촉연구원 CG/Vision연구실책임연구원 / 실장 ABSTRACT Recent technological advances in three-dimensional (3D) sensing devices and machine learning such as deep leaning has enabled data-driven 3D applications. Research on artificial intelligence has developed for the past few years and 3D deep learning has been introduced. This is the result of the availability of high-quality big data, increases in computing power, and development of new algorithms; before the introduction of 3D deep leaning, the main targets for deep learning were one-dimensional (1D) audio files and two-dimensional (2D) images. The research field of deep leaning has extended from discriminative models such as classification/segmentation/ reconstruction models to generative models such as those including style transfer and generation of non-existing data. Unlike 2D learning, it is not easy to acquire 3D learning data. Although low-cost 3D data acquisition sensors have become increasingly popular owing to advances in 3D vision technology, the generation/acquisition of 3D data is still very difficult. Even if 3D data can be acquired, post-processing remains a significant problem. Moreover, it is not easy to directly apply existing network models such as convolution networks owing to the various ways in which 3D data is represented. In this paper, we summarize technological trends in AI-based 3D content generation. KEYWORDS 딥러닝, 3D 콘텐츠, 3D 딥러닝, 3D 딥러닝표준 Ⅰ. 서론 Hand Crafted Feature 1 2014 2015 DOI https doi org 10 22648 ETRI 2019 J 340402 2018 R2018030391 3D 본저작물은공공누리제 4 유형출처표시 + 상업적이용금지 + 변경금지조건에따라이용할수있습니다. 2019 한국전자통신연구원
16 34 4 2019 8 5 4 5 Convolution 2 II 2 Hand Craft Feature Discriminative Model 1 X Y P Y X X Y Generative Model 2014 Ian J Goodfellow 3 GAN Generative Adversarial Network Y X 1 P X Y X Y Many to One Mapping One to Many Mapping GitHub Cumulative number of named GAN papers by month https deephunt in the gan zoo 79597dc8c347 3 2015 2017 2018 350 3D 1
17 3D 3D II 1. 학습데이터 1 2 3 3D 3D 2 2 a b c remesh 4 2 3 3 3D 2 a 3D ShapeNet Annotated 3D ShapeNetCor 55 51 300 ShapeNetStem Shape Net Core 12 000 https www shapenet org Human3 6M 6 5 17 3 6 DB http vision imar ro human3 6m ModelNet CAD 3D 662 127 915 http modelnet cs princeton edu CAESAR DB 5 http store sae org caesar 2 3D
18 34 4 2019 8 3 Ⅱ. 인공지능기반 3D 콘텐츠기술 1. 3D 데이터표현방법 2 3 5 6 4 2 RGB 1 6 3D 5 2 4 2D 5 vs
3D 19 6 3D 3D 7 RGB D 3D 360 3D 3D 3D 3D 3D Vertex 3D 2. 표현방법에따른학습방법 7 6 3D 5 6 3D 2D RGB D
20 34 4 2019 8 7 3D 3 3D 3D 2 8 GPU 8 2 3 9 3D 16 8 RGB skip connection latent code L1 L2 loss 9 8 3D 3D Eman 6
3D 21 8 3D 3D 3. 표준화동향 10 informative NNEF Neural Network Exchange Format https www khronos org nnef 1 0 TTA Telecommunication Technology Association PG 610 2D 3D 3D http tta or kr 3D 2D 3D 9 2D 3D XML Ⅲ. 결론 3D 3D 3D 3D
22 34 4 2019 8 용어해설 Hand Crafted Feature 기존영상인식등에서사용되는특징점. 예를들어, 얼굴인식의경우외곽선의모양비율등알고리즘에서정의한몇가지요소유클리디언평면고대의수학자유클리드가정의한평면으로유클리드의기하법칙 ( 예를들어, 임의의한점에서다른점으로직선을그을수있고, 직각은모두같다 ) 이적용되는공간 Latent Code 잠재변수로차원이축소된데이터특징벡터 [1] S. Khan and S.P. Yong, A Comparison of Deep Learning and Hand Crafted Features in Medical Image Modality Classification, in Proc. Int. Conf. Comput. Inform. Sci., Kuala Lumpur, Malaysia, Aug. 15-17, 2016, pp. 633-638. [2] Taewan. Kim, CNN, Convolution Neural Network 요약, Tawan.Kim Blog, Jan. 4, 2018, Available: http://taewan.kim/ post/cnn/ [3] I.J. Goodfellow et al., Generative Adversarial Nets. in Proc. Adv. Neural Inform. Process. Syst., Montreal, Canada, Dec. 8-13, 2014, pp. 1-9. [4] https://skymind.ai/wiki/open-datasets CNN GPU RGB D Convolutional Neural Network Graphics Processing Unit Red Green Blue Depth [5] 이승욱외, 3D 딥러닝기술동향, 전자통신동향분석제 33 권제 5 호, 2018,, pp. 103-110. [6] E. Ahmed et al., A survey on Deep Learning Advances on Different 3D Data Representations, 2018, arxive 1808.01462. [7] Z. Cao et al., 3D Object Classifcation via Spherical Projections, 2017, arxiv 1712.04426. [8] H. Su et al., Multi-view Convolutional Neural Networks for 3D Shape Recognition, Sep. 2015, arxiv 1505.00880. [9] 임성재외, 한장의 RGB 영상을이용한다시점뎁스맵생성기 술, 대한전자공학회 2019 년도하계종합학술대회, 2019. 6. [10] 강대기, 딥러닝을위한인공신경망표준포맷동향, TTA 저널, vol. 179, 2018, pp. 85-90.