Affinity Space Adaptation for Semantic Segmentation Across Domains

Zhou, Wei; Wang, Yukang; Chu, Jiajia; Yang, Jiehua; Bai, Xiang; Xu, Yongchao

doi:10.1109/tip.2020.3018221

Cited by 59 publications

(28 citation statements)

References 53 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…UDA aims to generalize the model learned from the labeled source domain to another unlabeled target domain. In the field of UDA, a group of approaches has shown promising results in object detection [9]- [22] and semantic segmentation [23]- [45], [53], [54]. The current mainstream approaches of these two tasks include adversarial learning [10]- [12], [26], [29], [55], selftraining [8], [41], [42], [56] and self-ensembling [39], [54], [57]- [63].…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Self-Adversarial Disentangling for Specific Domain Adaptation

Zhou¹,

Gu²,

Pang³

et al. 2021

Preprint

View full text Add to dashboard Cite

Domain adaptation aims to bridge the domain shifts between the source and target domains. These shifts may span different dimensions such as fog, rainfall, etc. However, recent methods typically do not consider explicit prior knowledge on a specific dimension, thus leading to less desired adaptation performance. In this paper, we study a practical setting called Specific Domain Adaptation (SDA) that aligns the source and target domains in a demanded-specific dimension. Within this setting, we observe the intra-domain gap induced by different domainness (i.e., numerical magnitudes of this dimension) is crucial when adapting to a specific domain. To address the problem, we propose a novel Self-Adversarial Disentangling (SAD) framework. In particular, given a specific dimension, we first enrich the source domain by introducing a domainness creator with providing additional supervisory signals. Guided by the created domainness, we design a self-adversarial regularizer and two loss functions to jointly disentangle the latent representations into domainness-specific and domainness-invariant features, thus mitigating the intra-domain gap. Our method can be easily taken as a plug-and-play framework and does not introduce any extra costs in the inference time. We achieve consistent improvements over state-of-the-art methods in both object detection and semantic segmentation tasks.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Annotating a large-scale dataset for each new domain is cost-expensive and timeconsuming. Unsupervised domain adaptation (UDA) emerges, which shows promising results on object detection [9]- [22] and semantic segmentation [23]- [45], aiming to reduce the domain shifts between the source and the target domains.…”

Section: Introductionmentioning

confidence: 99%

Self-Adversarial Disentangling for Specific Domain Adaptation

Zhou¹,

Gu²,

Pang³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Chang et al [11] construct the DISE framework which extracts domain-invariant structure and domain-specific texture information to reduce sourcetarget discrepancies. More recent works further prioritize category-level alignment (CLAN) [63], minimize adversarial entropy [67], or perform affinity-space domain adaptation [68].…”

Section: Unsupervised Domain Adaptive Semantic Segmentationmentioning

confidence: 99%

“…* denotes further adding source images from WildDash to complement Cityscapes. 68. 8.30 75.80 9.49 21.64 15.91 5.85 9.26 71.08 31.50 85.13 6.55 1.68 55.48 24.91 30.22 0.52 0.53 17.00 DANet S S 38.51 61.78 21.11 74.59 22.59 29.93 14.79 15.00 10.17 66.94 19.03 82.57 31.03 21.24 53.26 54.67 37.77 39.40 43.84 31.95 DANet A S 39.16 61.34 20.71 76.52 20.53 30.03 14.19 15.69 10.09 68.60 18.84 82.08 33.16 21.75 57.68 53.88 40.33 41.47 46.11 31.00 DANet S+A S 39.28 62.43 21.89 76.22 21.42 30.54 14.85 14.10 9.76 69.07 19.94 82.84 34.56 19.30 56.51 53.04 42.51 39.47 45.71 32.09 DANet S R 39.46 62.75 23.17 76.65 23.90 30.82 14.84 18.44 10.09 69.10 17.60 82.78 33.51 21.53 55.97 51.78 41.77 36.90 46.11 32.12 DANet S+A R 39.76 63.11 24.63 76.17 25.03 30.56 13.68 15.68 10.53 67.31 22.41 80.15 32.95 21.11 54.39 53.51 43.64 42.20 46.71 31.66 DANet S+A+F R 40.52 62.90 25.58 76.62 24.45 30.37 14.45 16.75 9.96 67.87 19.70 82.04 34.18 22.95 56.99 54.27 44.15 47.75 46.98 31.86 DANet-SSL S+A R 41.39 67.24 27.98 77.18 25.11 25.80 15.33 10.59 6.58 69.24 33.89 80.96 32.18 5.29 69.86 59.70 36.20 65.99 47.47 29.87 DANet-SSL S+A+F R 41.99 70.21 30.24 78.44 26.72 28.44 14.02 11.67 5.79 68.54 38.20 85.97 28.14 0.00 70.36 60.49 38.90 77.80 39.85 24.02 .63 35.30 78.52 25.27 33.51 14.43 13.80 7.31 63.52 34.94 84.31 34.54 19.08 70.05 49.14 48.80 75.11 47.53 35.36 DANet-SSL* S+A+F R 44.66 75.85 34.21 82.58 28.75 35.58 18.51 12.65 12.49 71.33 37.51 89.80 38.68 15.99 76.59 62.81 12.25 61.56 48.18 33.26…”

mentioning

confidence: 99%

Transfer beyond the Field of View: Dense Panoramic Semantic Segmentation via Unsupervised Domain Adaptation

Zhang¹,

Ma²,

Yang³

et al. 2021

Preprint

View full text Add to dashboard Cite

“…Pixels neighborhood relationship approach was recently studied to solve the above problem. Zhou [8] presented an affinity space for semantic segmentation International Journal of Intelligent Engineering and Systems, Vol. 15 that highlights structure using co-occurring output patterns between neighboring pixels.…”

Section: Introductionmentioning

confidence: 99%

Improvement for Single Image Super-resolution and Image Segmentation by Graph Laplacian Regularizer Based on Differences of Neighboring Pixels

2022

IJIES

View full text Add to dashboard Cite

The neighboring relationship of the pixels is fundamental information in the image. Utilizing neighboring pixel information provides an advantage over using only pixel-to-pixel information. This study presents a regularization term based on the differences of neighboring pixels. We define the differences of neighboring pixels by using the Graph Theory approach. We then propose an effective graph Laplacian regularization to enforce the differences between pixels as the nodes and differences between nodes as edges. In our scheme, Graph Laplacian constructs from the output of the convolutional neural network and ground-truth image. The generated differences matrices from the output and ground-truth are combined as a Laplacian regularization term used as the deep convolutional neural network's new objective function. Experiments show that our scheme successfully captures pixel neighbor relations and improves the performance of the convolutional neural network better than the baseline without a regularization term. Qualitative and quantitative results show that our developed regularizer proved to enhance the boundary structure on image super-resolution and image segmentation tasks, which is achieved Structural Similarity Index (SSIM) of 0.9604 and 0.8263 on Cartoon set and Manga109 datasets and area under the curve (AUC) of 0.9740 and 0.9561 on DRIVE and STARE datasets, respectively.

show abstract

Affinity Space Adaptation for Semantic Segmentation Across Domains

Cited by 59 publications

References 53 publications

Self-Adversarial Disentangling for Specific Domain Adaptation

Self-Adversarial Disentangling for Specific Domain Adaptation

Transfer beyond the Field of View: Dense Panoramic Semantic Segmentation via Unsupervised Domain Adaptation

Improvement for Single Image Super-resolution and Image Segmentation by Graph Laplacian Regularizer Based on Differences of Neighboring Pixels

Contact Info

Product

Resources

About