Open Set Domain Adaptation by Backpropagation

By Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, Tatsuya Harada. European Conference on Computer Vision, 2018

610 citations recorded by the source corpus.

Numerous algorithms have been proposed for transferring knowledge from a label-rich domain (source) to a label-scarce domain (target). Most of them are proposed for closed-set scenario, where the source and the target domain completely share the class of their samples. However, in practice, a target domain can contain samples of classes that are not shared by the source domain. We call such classes the “unknown class” and algorithms that work well in the open set situation are very practical. However, most existing distribution matching methods for domain adaptation do not work well in this setting because unknown target samples should not be aligned with the source. In this paper, we propose a method for an open set domain adaptation scenario, which utilizes adversarial training. This approach allows to extract features that separate unknown target from known target samples. During training, we assign two options to the feature generator: aligning target samples with source known ones or rejecting them as unknown target ones. Our method was extensively evaluated and outperformed other methods with a large margin in most settings.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/cf6e4e030b50ca7815dbdd6bf181f4b51e6a2c4c

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 Computer Vision and Pattern Recognition

Citations in the corpus

Confidence Regularized Self-Training (2019, background): “Open set domain adaptation [40, 52] focuses on the problem where classes are not totally shared between source and target domains.”
A Brief Review of Domain Adaptation (2020, background): “[21], introduced new open set domain adaptation in which the data in the source private classes is removed.”
Separate to Adapt: Open Set Domain Adaptation via Progressive Separation (2019, methodology): “We follow previous work [30] using the same set of known classes and unknown classes in the target domain.”
Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-Identification (2019, background): “Recently, an adversarial learning framework [22] is proposed to separate target samples into known and unknown classes, and reject unknown classes during feature alignment.”
Universal Source-Free Domain Adaptation (2020, background): “In the second kind, Open-set DA [39, 1, 10], the target label space is considered as a superset of the source label space (i.”
Learning to Transfer Examples for Partial Domain Adaptation (2019, background): “Unlike OSDA, in our scenario the target classes are entirely unknown at training.”
Class-Incremental Domain Adaptation (2020, methodology): “We compare against prior UDA methods DANN [11], OSBP [45], UAN [54], STA [29], and CI methods E2E [4], LETR [33], iCaRL [41], LwF-MC [7], LwM [7].”
Minimum Class Confusion for Versatile Domain Adaptation (2019, background): “ng performance and accelerating training. MCC can be applied directly to all of the DA scenarios where there are no unknown classes in the target domain. The variants of MCC which can tackle Open-Set [2,44,24] or Universal [58] scenarios are remained to be explored. Moreover, we can consider extendi”
FSDR: Frequency Space Domain Randomization for Domain Generalization (2021, background): “such models usually undergo a drastic performance drop while applied to real-world images [78] due to the domain bias and shift [55, 41, 67, 71, 56].”
Domain Consensus Clustering for Universal Domain Adaptation (2021, methodology): “OSBP [54] proposes an adversarial learning framework that enables the feature generator to learn representations to achieve common-private separation.”
Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation (2020, methodology): “For speciﬁc partial-set and open-set DA tasks, we compare with IWAN (Zhang et al., 2018a), SAN (Cao et al., 2018), ETN (Cao et al., 2019), SAFN, and ATI-λ (Panareda Busto & Gall, 2017), OSBP (Saito et al., 2018c), OpenMax (Bendale & Boult, 2016), STA (Liu et al., 2019), respectively.”
Recent Advances in Open Set Recognition: A Survey (2018, background): “For example, [156], [157], [158] have introduced this scenario into domain adaptation, while [159] introduced it to the semantic instance segmentation task.”
Under review as a conference paper at ICLR 2020 many domain adaptation methods (2019, methodology): “The averaged accuracy over the classes Saito et al. (2018); Tan et al. (2019) is used to OpenSet1 and OpenSet2, and for other experiments the classification accuracy is used as the metric.”
Applications of Unsupervised Deep Transfer Learning to Intelligent Fault Diagnosis: A Survey and Comparative Study (2019, methodology): “Following previous work in [144], [150], [176], [177], five evaluation metrics are employed: 1) Accuracy of shared classes: ALL∗ = MS NS 2) Accuracy of unknown classes: UNK = MU NU 3) Averaged accuracy of all classes: OS = 1 C+1 ∑C c=0Accc 4) Accuracy of all test samples: ALL = MS+MU NS+NU 5) Harmon”
Universal Domain Adaptation through Self Supervision (2020, background): “Open set Domain Adaptation (ODA) deals with target examples whose class is different from any of the source classes [28, 35, 23].”
Open Compound Domain Adaptation (2019, background): “Another route of research aims to tackle the category sharing/unsharing issues between source and target domain, namely open set [37, 43] and partial set [55, 3] domain adaptation.”