Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation

By Yawei Luo, Liang Zheng, Tao Guan, Junqing Yu, Yi Yang. 2018

817 citations recorded by the source corpus.

We consider the problem of unsupervised domain adaptation in semantic segmentation. The key in this campaign consists in reducing the domain shift, i.e., enforcing the data distributions of the two domains to be similar. A popular strategy is to align the marginal distribution in the feature space through adversarial learning. However, this global alignment strategy does not consider the local category-level feature distribution. A possible consequence of the global movement is that some categories which are originally well aligned between the source and target may be incorrectly mapped. To address this problem, this paper introduces a category-level adversarial network, aiming to enforce local semantic consistency during the trend of global alignment. Our idea is to take a close look at the category-level data distribution and align each class with an adaptive adversarial loss. Specifically, we reduce the weight of the adversarial loss for category-level aligned features while increasing the adversarial force for those poorly aligned. In this process, we decide how well a feature is category-level aligned between source and target by a co-training approach. In two domain adaptation tasks, i.e., GTA5 -> Cityscapes and SYNTHIA -> Cityscapes, we validate that the proposed method matches the state of the art in segmentation accuracy.

Subfields: Artificial Intelligence, Radiology, Nuclear Medicine and Imaging

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

Unsupervised Model Adaptation for Continual Semantic Segmentation (2020, background): “As a result, data annotation is a major bottleneck to address the problem of domain shift, where domain discrepancy exists between the distributions of training and testing domains (Luo et al. 2019) and the trained model needs to be adapted to generalize again after being fielded.”
Contrastive Adaptation Network for Unsupervised Domain Adaptation (2019, background): “Recent advancements in deep neural networks have successfully improved a variety of learning problems [40, 8, 26, 19, 20].”
Domain Generalization via Model-Agnostic Learning of Semantic Features (2019, methodology): “This is practically performed by narrowing the domain shift between the target and source either in input space [19], feature space [6, 11, 32, 42, 51], or output space [33, 43, 49], generally using maximum mean discrepancy [15, 46] or adversarial learning [14].”
Pose-Guided Feature Alignment for Occluded Person Re-Identification (2019, background): “Recent years, deep learning methods [39, 34, 13, 27, 3, 25, 7, 6, 36, 21] for person re-id show signiﬁcant superiority on retrieval accuracy.”
Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration (2018, background): “The deeper and wider architectures of deep CNNs bring about the superior performance of computer vision tasks [6, 26, 45].”
OpenEarthMap: A Benchmark Dataset for Global High-Resolution Land Cover Mapping (2022, methodology): “For the unsupervised domain adaptation task, a metricbased method (MCD [46]), adversarial training methods including AdaptSeg [45], category-level adversarial network (CLAN) [28], TransNorm [51], and fine-grained adversarial learning framework for domain adaptive (FADA) [48]), as well as self-traini”
DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation (2021, background): “Using multiple scales [9, 72] or category information [16, 50, 83] for the discriminator can refine the alignment.”
ACDC: The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding (2021, background): “A major class of UDA works employs adversarial domain adaptation to implicitly align the source and target domains at the level of pixels and/or features [7, 15, 26, 41, 43, 44, 46, 48, 60].”
BAPA-Net: Boundary Adaptation and Prototype Alignment for Cross-domain Semantic Segmentation (2021, methodology): “Following the common experimental setup for crossdomain semantic segmentation [50, 37, 38, 51, 52, 61, 66, 64], we evaluate our method on two synthetic-to-real scenarios, namely GTA5 to Cityscapes and SYNTHIA to Cityscapes.”
Rectifying Pseudo Label Learning via Uncertainty Estimation for Domain Adaptive Semantic Segmentation (2020, methodology): “Luo et al. [Luo et al., 2019a,Luo et al., 2019b] utilize the attention mechanism to reﬁne the feature alignment.”
Cross-Domain Gradient Discrepancy Minimization for Unsupervised Domain Adaptation (2021, methodology): “Recently, CLAN [28] use cosine similarity of the classifier parameters to measure the discrepancy for semantic segmentation.”
Discover, Hallucinate, and Adapt: Open Compound Domain Adaptation for Semantic Segmentation (2021, background): “In particular, existing UDA techniques mostly focus on a single-source single-target setting [37, 39, 45, 13, 25, 31, 5, 29].”
Learning Texture Invariant Representation for Domain Adaptation of Semantic Segmentation (2020, result): “Therefore, we set smaller value than prior works which adopt the same architecture [23, 18, 25, 1, 16].”
Stochastic Classifiers for Unsupervised Domain Adaptation (2020, methodology): “Similarly, MCD [45] adapted the target features by aligning the outputs of diverse classifiers and CLAN [31] designed a category-level adversarial loss for semantic segmentation.”
Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-Identification (2019, background): “The art supervised methods have made great achievement in person re-ID [14, 26, 39, 25], relying on rich-labeled data and the success of deep networks [18, 12, 8].”