High-Resolution Deep Image Matting

Image matting is a key technique for image and video editing composition. Conventionally, deep learning approaches take the whole input an associated trimap to infer alpha matte using convolutional neural networks. Such set state-of-the-arts in matting; however, they may fail real-world applications due hardware limitations, since images are mostly of very high resolution. In this paper, we propose HDMatt, first based approach high-resolution inputs. More concretely, HDMatt runs patch-based crop-and-stitch manner inputs with novel module design address contextual dependency consistency issues between different patches. Compared vanilla inference which computes each patch independently, explicitly model cross-patch newly-proposed Cross-Patch Contextual (CPC) guided by given trimap. Extensive experiments demonstrate effectiveness proposed method its necessity Our also sets new state-of-the-art performance on Adobe Matting AlphaMatting benchmarks produce impressive visual results more images.