AMICO: Amodal Instance Composition

BMVC 2021

Peiye Zhuang1,3,    Jia-Bin Huang2,3,    Ayush Saraf3,    Xuejian Rong3,    Changil Kim3,    Denis Demandolx3,   

1University of Illinois, Urbana-Champaign   
2University of Maryland, College Park 3Meta

Image composition aims to blend multiple objects to form a harmonized image. Existing approaches often assume precisely segmented and intact objects. Such assumptions, however, are hard to satisfy in unconstrained scenarios. We present Amodal Instance Composition for compositing imperfect -- potentially incomplete and/or coarsely segmented -- objects onto a target image. We first develop object shape prediction and content completion modules to synthesize the amodal contents. We then propose a neural composition model to blend the objects seamlessly. Our primary technical novelty lies in using separate foreground/background representations and blending mask prediction to alleviate segmentation errors. Our results show state-of-the-art performance on public COCOA and KINS benchmarks and attain favorable visual results across diverse scenes. We demonstrate various image composition applications such as object insertion and de-occlusion.


Composition with imperfect instances. Our method takes an ordered collection of imperfect instances (i.e., partially occluded and/or coarsely cropped) from multiple source images (instances identified by colored bounding boxes) and a background image as inputs (a), and produces harmonized composition (b). We achieve this via a unified framework that estimates the precise shape and content of each object and adjusts the object appearances to be mutually compatible.



<1> Object re-shuffling

(left: original images. The arrows indicate object moving directions; right: results with objects re-shuffled.)

<2> Object insertion

(We insert the dishes into the background image. Our method harmonizes the content and refines the imperfect masks.)


				title = {AMICO: Amodal Instance Composition},
				author = {Zhuang, Peiye and 
				Huang, Jia-bin and 
				Saraf, Ayush and 
				Rong, Xuejian and 
				Kim, Changil and 
				Demandolx, Denis},
				booktitle = {Proc. BMVC},
				year = {2021},


Some of the work was completed while P.Z. was at Meta.