Title:
TRAINING DATA GENERATION METHOD AND GENERATION APPARATUS, AND IMAGE SEMANTICS SEGMENTATION METHOD THEREFOR
Document Type and Number:
WIPO Patent Application WO/2019/019019
Kind Code:
A1
Abstract:
A training data generation method and generation apparatus, and an image semantics segmentation method therefor, the training data generation method comprising: setting a corresponding category label for an object model in a three-dimensional stereoscopic scene (201); acquiring several corresponding scene images by changing scene parameters of the three-dimensional stereoscopic scene (202); rendering the object model as a monochrome material corresponding to the category label of the object model (203); acquiring a semantics segmentation image corresponding to each scene image based on the rendered object model (204); and storing each scene image and the semantics segmentation image corresponding to each scene image as a set of training data (205). The method can quickly generate a large number of scene images and semantic segmentation images that meet the usage requirements and reflect different times, thereby greatly saving time and human labour.
More Like This:
Inventors:
WANG KAI (CN)
LIAN SHIGUO (CN)
WANG LUOWEI (CN)
LIAN SHIGUO (CN)
WANG LUOWEI (CN)
Application Number:
PCT/CN2017/094312
Publication Date:
January 31, 2019
Filing Date:
July 25, 2017
Export Citation:
Assignee:
CLOUDMINDS SHENZHEN ROBOTICS SYSTEMS CO LTD (CN)
International Classes:
G06T7/50; G06V10/772; G06V10/774; G06V20/00; G06V20/70
Foreign References:
| CN106169082A | 2016-11-30 | |||
| CN103268635A | 2013-08-28 | |||
| EP3156942A1 | 2017-04-19 | |||
| US20170193340A1 | 2017-07-06 |
Other References:
See also references of EP 3660787A4
Attorney, Agent or Firm:
SHENZHEN AIDISEN INTELLECTUAL PROPERTY OFFICE (CN)
Download PDF: