Disentangled Image Matting

Cai, Shaofan; Zhang, Xiaoshuai; Fan, Haoqiang; Huang, Haibin; Liu, Jiangyu; Liu, Jiaming; Liu, Jiaying; Wang, Jue; Sun, Jian

doi:10.48550/arxiv.1909.04686

Cited by 5 publications

(7 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our method achieves new state-of-the-arts on AIM, AlphaMatting benchmarks and produce impressive visual results on real-world high-resolution images. [1], SampleNet [35], GCA Matting [24] and HDMatt (Ours).…”

Section: Discussionmentioning

confidence: 99%

“…On the first stage, we pre-trained a Resnet-34 classification model on ImageNet [10]. We follow the same training configuration as the public PyTorch implementation 1 . Then all layers from the model before the fully-connected layer were used as our matting encoder.…”

Section: Methodsmentioning

confidence: 99%

“…Given an input image and its trimap indicating the background, foreground and unknown regions, image matting is applied to estimate the alpha matte inside the unknown region to clearly separate the foreground from the background. Recently, many deeplearning-based methods [39,26,17,1,34] have achieved significant improvements over traditional methods [37,13,33,23,14]. These deep learning methods mostly take the whole images and the associated whole trimaps as the inputs, and employ deep neural networks such as VGG [32] Figure 1: Down-sampling (DS) and cropping (C) strategies applied to ContextNet [17] on an HR image.…”

Section: Introductionmentioning

confidence: 99%

“…Figure 8: Visual results on alphamatting.com testset. From left to right: Image, Trimap (U), AdaMatting[1], SampleNet[35], GCA Matting[24] and HDMatt (Ours).…”

mentioning

confidence: 99%

See 3 more Smart Citations

High-Resolution Deep Image Matting

Yu¹,

Xu²,

Huang³

et al. 2020

Preprint

View full text Add to dashboard Cite

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

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

“…Figure 8: Visual results on alphamatting.com testset. From left to right: Image, Trimap (U), AdaMatting[1], SampleNet[35], GCA Matting[24] and HDMatt (Ours).…”

mentioning

confidence: 99%

See 2 more Smart Citations

High-Resolution Deep Image Matting

Yu¹,

Xu²,

Huang³

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…Given an input image and a trimap indicating the background, forground and unknown regions, image matting is applied to estimate the alpha matte inside the unknown region to clearly separate the foreground from the background. Recently, many deep-learning-based methods (Xu et al 2017;Lu et al 2019;Hou and Liu 2019;Cai et al 2019) have achieved significant improvements over traditional methods (Wang and Cohen 2007;Gastal and Oliveira 2010;Sun et al 2004;Levin, Lischinski, and Weiss 2007;Grady et al 2005). These deep learning methods (Xu et al 2017;Lu et al 2019;Hou and Liu 2019) mostly take the whole images and the associated whole trimaps as the inputs, and employ deep neural networks such as VGG (Simonyan and Zisserman 2014) and Xception (Chollet 2017) as their network backbones.…”

Section: Introductionmentioning

confidence: 99%

High-Resolution Deep Image Matting

Huang

et al. 2021

AAAI

View full text Add to dashboard Cite

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

show abstract

MolexNet: Real-Time Recurrent-Based Object Removal and Enhancement Network

Wang,

Wang

2024

Preprint

View full text Add to dashboard Cite

In the contemporary realm of human portrait image segmentation tasks, achieving precise separation of foreground and background within intricate background and foreground edge scenarios poses a formidable challenge. In this study, we introduce a robust, efficient, and remarkably accurate image segmentation model named "MolexNet". In the domain of human portrait image segmentation, our model outperforms the majority of network architecture models in the field. What sets our model apart from conventional lightweight segmentation models is the clever fusion of features at different scales from the input image within the decoder. This enhancement significantly enhances processing speed while maintaining the quality of image segmentation, resulting in exceptional results for human portrait image segmentation tasks. We rigorously evaluate this model using various test datasets, and it consistently exhibits outstanding performance in both image and video segmentation tasks. It adeptly handles complex scenes and dynamic foreground elements, yielding satisfying segmentation results. Moreover, due to its ingenious design concepts, precise optimization strategies, and efficient model structure, our model maintains a high level of accuracy. We also conduct comparative analyses with other open-source architectures to underscore the superior performance and computational efficiency of our model. The model presented in this paper serves as a valuable reference and source of inspiration for research and practical applications in the field of image segmentation.

show abstract

Disentangled Image Matting

Cited by 5 publications

References 33 publications

High-Resolution Deep Image Matting

High-Resolution Deep Image Matting

High-Resolution Deep Image Matting

MolexNet: Real-Time Recurrent-Based Object Removal and Enhancement Network

Contact Info

Product

Resources

About