Video -synthesized NVIDIA etc. according to the flaky paper, etc. Technical development: Innovative Tech

Innovative Tech:

In this corner, Hiroki Yamashita, who presides over the web media "SEAMLESS", introducing the latest research on technology.Mr. Yamashita picks up and explains a highly new nature paper.

　"Life: Lighting Invariant Flow Estimation" developed by a research team by Hong Kong Junbun University, Zhejiang University, and NVIDIA in the United States, an optical flow estimation that allows you to convert images in videos into different images and images (each pixel of the moving objectsIt is a deep learning framework for the amount of mobility of each pixel that represents which position has been moved in numerical values.

　The image can be converted stably even if the camera's viewpoint and lighting are very different.In addition, it is possible to synthesize as if it is sticky, without shifting or disturbing the surface like a paper that bends complicatedly.

　Even if the camera perspective is moved, even if the camera perspective is moved, even if the camera perspective is moved, even if the camera perspective is moved, the moving object is estimated by machine learning.It is known that the amount of matching method is highly accurate.

　However, if the camera's viewpoint changes greatly, the accuracy will be significantly reduced if the change in lighting is large.Images are disturbed in the case of paper with complex deformation.

　In order to challenge these issues, the research team has built a weak religion for an optical flow that can convert images stably even if the camera perspective or lighting fluctuations are large.Estimate the relative camera pose between images and match the characteristics between images.

　In order to evaluate the effectiveness, an experiment was conducted to convert other images into paintings printed on A4 -sized paper.As a result of comparing with other similar methods, it has achieved more stable conversion than any method.You can check the comparative video on the project page.

　This method can also adapt to a variety of applications that require a relative posture, visual position identification, homolog conversion, and AR (expanding reality).

　Click here for the video.

関連記事

関連リンク

To read more, you need to agree to the terms of use and register "ITMEDIA NEWS Anchor Desk Magazine".