A new method for image composition of partially occluded objects is proposed by the group led by Prof. WANG Wencheng from the state key laboratory of computer science, Institute of Software, Chinese Academy of Sciences. This work was presented in Pacific Graphics 2019 and published in Computer Graphics Forum, Vol.37, No.7, 2019， entitled " Image composition of partially occluded objects".

Image composition tries to blend some contents of the source image into the target image seamlessly. Existing methods always require the contents from the source image must be placed over the contents of the target image. Thus, they are very difficult to blend the contents that are occluded partially by the contents of the target image, which always requires the user to interactively remove the occlude portions, or carefully order the contents before composition. This seriously prevents image composition efficiency.

We observe that the contents from the source image must be presented enough in the composited image, and in image composition, the contents from the source image are always placed to be on a plane that are shared by many contents of the target image. Based on the observation, we develop novel measures that can well deduce the occlusion relationships between the contents from the source image and the contents of the target image. Thus, we propose a new method that can efficiently perform image composition of partially occluded objects. At first, we employ learning-based methods to segment the objects in the source image and the target image, and obtain the depth information of the target image. Then, we allow the user to select contents from the source image by simple interactivities, such as only clicking his interested objects. After the user places his selected contents into the target image, our method can well deduce the corresponding occlusion relationships between the contents from the source image and the contents of the target image, and so perform image composition quickly. As the new method avoids the limit that the contents from the source image must be placed over the contents of the target image, it considerably promotes image composition. This is very helpful for applications.

The work is partially supported by the national natural science foundation of China.