Reducing Domain Gap via Style-Agnostic Networks
By Hyeonseob Nam, Hyunjae Lee, Jongchan Park, Wonjun Yoon, Donggeun Yoo. 2019
66 citations recorded by the source corpus.
Deep learning models often fail to maintain their performance on new test domains. This problem has been regarded as a critical limitation of deep learning for real-world applications. One of the main causes of this vulnerability to domain changes is that the model tends to be biased to image styles (i.e. textures). To tackle this problem, we propose Style-Agnostic Networks (SagNets) to encourage the model to focus more on image contents (i.e. shapes) shared across domains but ignore image styles. SagNets consist of three novel techniques: style adversarial learning, style blending and style consistency learning, each of which prevents the model from making decisions based upon style information. In collaboration with a few additional training techniques and an ensemble of several model variants, the proposed method won the 1st place in the semi-supervised domain adaptation task of the Visual Domain Adaptation 2019 (VisDA-2019) Challenge.
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
- Divyat Mahajan
- David Lopez-Paz
- Daehee Kim
- Shruti Tople
- Seunghyun Park
- Jinkyu Kim
- Amit Sharma
- Ishaan Gulrajani
Citations in the corpus
- In Search of Lost Domain Generalization (2020, background): “Nam et al. [2019], Asadi et al. [2019] promote representations of data that ignore texture and focus on shape.”
- SelfReg: Self-supervised Contrastive Regularization for Domain Generalization (2021, background): “This testbed currently provides seven multi-domain datasets (i.e. ColoredMNIST [1], RotatedMNIST [16], VLCS [13], PACS [26], Office-Home [42], and TerraIncognita [2], DomainNet [34]) and provides benchmarks results of 14 baseline approaches (i.e. ERM [41], IRM [1], GroupDRO [36], Mixup [45], MLDG [27”
- Domain Generalization using Causal Matching (2020, methodology): “On the PACS dataset with ResNet-18 architecture (Table 4), our methods are competitive to state-of-the-Domain (Carlucci et al., 2019), DDAIG (Zhou et al., 2020), SagNet (Nam et al., 2019), DDEC (Asadi et al., 2019), were taken from the Do-mainBed (Gulrajani & Lopez-Paz, 2020) paper.”