Location-aware Single Image Reflection Removal

Dong, Zheng; Yang, Yin; Bao, Hujun; Xu, Weiwei; Lau, Rynson W. H.

doi:10.1109/iccv48922.2021.00497

Cited by 70 publications

(19 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To alleviate the misalignment problem, Wei et al [49] leveraged the highest-level VGG [37] features which are insensitive to misalignment. Reflection distribution is also considered to facilitate SIRR, Dong et al [6] predicted a reflection confidence map which is integrated into a composition loss, while Li et al [25] defined an explicit mask loss by leveraging reflection strength. In this paper, we enhance the SIRR task from a domain generalization perspective, which is orthogonal to and cooperates well with the above methods.…”

Section: Learning-based Sirr Methodsmentioning

confidence: 99%

“…CEILNet presented by Fan et al [9] is a two-stage architecture, where the edge maps are predicted before the estimation of the transmission layer. The two-stage setting is followed and developed by DMGN [10], Dong et al [6], and RAGNet [25], and they estimate the reflection layer in the prior stage instead. Zhang et al [55] stacked the features extracted by multiple layers of a pre-trained VGG-19 [37] model at the beginning of SIRR models, which is retained in a series of subsequent works [10,49].…”

Section: Learning-based Sirr Methodsmentioning

confidence: 99%

“…Zhang et al [55] stacked the features extracted by multiple layers of a pre-trained VGG-19 [37] model at the beginning of SIRR models, which is retained in a series of subsequent works [10,49]. Yang et al [53] proposed to predict the reflection layer and the transmission layer iteratively, and IBCLN [21] further introduced recurrent neural networks, which is also exploited in [6,10]. Prasad et al [31] proposed a multi-scale structure which starts from a low-resolution input and progressively grows into the desired resolution.…”

Section: Learning-based Sirr Methodsmentioning

confidence: 99%

“…When capturing images through glass, undesired reflection by the glass surface is one of the key factors causing image quality degradation, and the existence of these reflections also hampers many other vision applications such as image classification and autonomous driving. In recent years, driven by the development of deep neural networks (DNNs), learning-based methods have achieved continuous improvements on single-image reflection removal (SIRR) tasks [6,9,10,16,21,25,31,44,49,50,53,55,56].…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Adaptive Network Combination for Single-Image Reflection Removal: A Domain Generalization Perspective

Liu¹,

Pan²,

Yan³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recently, multiple synthetic and real-world datasets have been built to facilitate the training of deep single image reflection removal (SIRR) models. Meanwhile, diverse testing sets are also provided with different types of reflection and scenes. However, the non-negligible domain gaps between training and testing sets make it difficult to learn deep models generalizing well to testing images. The diversity of reflections and scenes further makes it a mission impossible to learn a single model being effective to all testing sets and real-world reflections. In this paper, we tackle these issues by learning SIRR models from a domain generalization perspective. Particularly, for each source set, a specific SIRR model is trained to serve as a domain expert of relevant reflection types. For a given reflectioncontaminated image, we present a reflection type-aware weighting (RTAW) module to predict expert-wise weights. RTAW can then be incorporated with adaptive network combination (AdaNEC) for handling different reflection types and scenes, i.e., generalizing to unknown domains. Two representative AdaNEC methods, i.e., output fusion (OF) and network interpolation (NI), are provided by considering both adaptation levels and efficiency. For images from one source set, we train RTAW to only predict expert-wise weights of other domain experts for improving generalization ability, while the weights of all experts are predicted and employed during testing. An in-domain expert (IDE) loss is presented for training RTAW. Extensive experiments show the appealing performance gain of our AdaNEC on different state-of-the-art SIRR networks. Source code and pre-trained models will available at https://github. com/csmliu/AdaNEC.

show abstract

Section: Learning-based Sirr Methodsmentioning

confidence: 99%

Section: Learning-based Sirr Methodsmentioning

confidence: 99%

Section: Learning-based Sirr Methodsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Adaptive Network Combination for Single-Image Reflection Removal: A Domain Generalization Perspective

Liu¹,

Pan²,

Yan³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Classical methods use image priors to solve this illposed problem, including gradient sparsity [Arvanitopoulos et al 2017;Levin and Weiss 2007], smoothness priors [Li and Brown 2014;Wan et al 2016], and ghosting cues [Shih et al 2015]. Recently, deep learning has been used for this task [Dong et al 2021;Fan et al 2017;Hong et al 2021;Li et al 2020b;Wan et al 2018; Our task is in-between these classes, because the reflections are on the floor rather than on glass surfaces but the appearance is similar to that of glass reflection because the floor is flat. To our knowledge, this scenario has not been well studied.…”

Section: Related Workmentioning

confidence: 99%