Netmarble announces Deep Learning Model recognizing key objects in detail
Netmarble announces Deep Learning Model recognizing key objects in detail
  • Dan Yoo
  • 승인 2018.10.10 13:54
  • 댓글 0
이 기사를 공유합니다

Netmarble announced on October 10 that Netmarble AI Revolution Center (NARC) and European Conference on Computer Vision (CCV) won second place at COCO DensePose Challenge. The top spot went to Beijing Post University in China.  

COCO DensePose Challenge, which was held in Muchen, Germany in last September, is a competition that announces technologies that uses deep learning to recognize images of people and map them to 3D images of the human body for poses. (DensePose)

In particular, the challenge focused on instance segmentation, human key point recognition, and detection of objects using DensePose. 

Netmarble had a good result by announcing a model in which Deep Learning recognizes major objects more accurately and in detail. 

Netmarble decided that it was difficult to detect a person with a smaller size due to low resolution of Feature Map during the process of DensePose's recognition and prediction. 

To solve this problem, Netmarble composed feature maps of different layers of different resolution and stacked the Up-Sampling layer tightly. Up-Sampling is the conversion of low-resolution images to high-resolution images by increasing pixels.

"The technology can be used throughout game art by modeling character pose in 2D images and applying texture," said Ahn Soo-nam, head of NARC game development AI team who participated in this challenge.

삭제한 댓글은 다시 복구할 수 없습니다.
그래도 삭제하시겠습니까?
댓글 0
계정을 선택하시면 로그인·계정인증을 통해
댓글을 남기실 수 있습니다.

  • Korea IT Times: Copyright(C) 2004, Korea IT Times. .Allrights reserved.
  • #1206, 36-4 Yeouido-dong, Yeongdeungpo-gu, Seoul, Korea(Postal Code 07331)
  • 서울특별시 영등포구 여의도동 36-4 (국제금융로8길 34) / 오륜빌딩 1206호
  • * Mobile News:
  • * Internet news:
  • * Editorial Div. 02-578-0434 / 010-2442-9446 * PR Global/AD: 82-2-578-0678.
  • * IT Times Canada: Willow St. Vancouver BC
  • 070-7008-0005
  • * Email: