Foreground-aware Semantic Representations for Image Harmonization

Sofiiuk, Konstantin; Popenova, Polina; Konushin, Anton

doi:10.1109/wacv48630.2021.00166

Cited by 57 publications

(53 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Recently, thanks to the inspiring data acquisition approach [37], deep learning based methods spring up rapidly. In [30,37], they both leveraged auxiliary semantic features to improve the basic image harmonization network. In [6], Cong released the first large-scale image harmonization dataset iHarmony4 and introduced a domain verification discriminator pulling close the foreground domain and background domain.…”

Section: Image Harmonizationmentioning

confidence: 99%

“…The pipeline of our network is shown in Figure 3. The generator G adopts the improved backbone proposed in iDIH [30], where the image blending layer is added to the UNet-like architecture. Since both input space and output space are different across two domains, we split the first (resp., last) 4 layers in the encoder (resp., decoder) into E rd (resp., D rd ) and E rl (resp., D rl ), where E rd (resp., E rl ) and D rd (resp., D rl ) are the domain-specific encoder and decoder for the rendered (resp., real) image domain.…”

Section: Cross-domain Harmonization Networkmentioning

confidence: 99%

Section: Implementation Detailsmentioning

confidence: 99%

“…We adopt the UNet-like architecture of iDIH [30] as backbone considering its simplicity and effectiveness. Note that we do not use auxiliary semantic information as in [30] for brevity and fair comparison with other methods. As in Figure 3, the domain-specific encoder (resp., decoder) contains the first (resp., last) 4 layers in the encoder (resp., decoder).…”

Section: Implementation Detailsmentioning

confidence: 99%

“…Image harmonization aims to diminish the discrepancy by adjusting the appearance of the composite foreground according to the background. Recently, many deep learning based harmonization methods [6,7,16,30,37] have achieved promising results.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Deep Image Harmonization by Bridging the Reality Gap

Cong

Cao

Niu

et al. 2021

Preprint

View full text Add to dashboard Cite

Image harmonization has been significantly advanced with large-scale harmonization dataset. However, the current way to build dataset is still labor-intensive, which adversely affects the extendability of dataset. To address this problem, we propose to construct a large-scale rendered harmonization dataset RHHarmony with fewer human efforts to augment the existing real-world dataset. To leverage both real-world images and rendered images, we propose a Cross-domain harmonization network CharmNet to bridge the domain gap between two domains. Moreover, we also employ well-designed style classifiers and losses to facilitate cross-domain knowledge transfer. Extensive experiments demonstrate the potential of using rendered images for image harmonization and the effectiveness of our proposed network.

show abstract