Adversarial Network With Multiple Classifiers for Open Set Domain Adaptation

Abstract

Domain adaptation aims to transfer knowledge from a domain with adequate labeled samples to a domain with scarce labeled samples. Prior research has introduced various open set domain adaptation settings in the literature to extend the applications of domain adaptation methods in realworld scenarios. This paper focuses on the type of open set domain adaptation setting where the target domain has both private ('unknown classes') label space and the shared ('known classes') label space. However, the source domain only has the 'known classes' label space. Prevalent distribution-matching domain adaptation methods are inadequate in such a setting that demands adaptation from a smaller source domain to a larger and diverse target domain with more classes. For addressing this specific open set domain adaptation setting, prior research introduces a domain adversarial model that uses a fixed threshold for distinguishing known from unknown target samples and lacks at handling negative transfers. We extend their adversarial model and propose a novel adversarial domain adaptation model with multiple auxiliary classifiers. The proposed multi-classifier structure introduces a weighting module that evaluates distinctive domain characteristics for assigning the target samples with weights which are more representative to whether they are likely to belong to the known and unknown classes to encourage positive transfers during adversarial training and simultaneously reduces the domain gap between the shared classes of the source and target domains. A thorough experimental investigation shows that our proposed method outperforms existing domain adaptation methods on a number of domain adaptation datasets. Index Terms-Open set domain adaptation, adversarial domain networks, multi-classifier based weighting module. I. INTRODUCTIOND EEP learning models for computer vision tasks usually require a massive amount of labeled data entailing highly laborious work for annotating data [1], [2], [3], [4]. An alternative is to use labeled data from a related (source) domain to boost the performance of the model in a target domain. However, as the source and target data may have domain gaps such as different illumination set-ups, and perspectives, synthesized data by using different variants of sensors, the performance of this approach may suffer. Existing domain adaptation (DA) methods aim to decrease the above-mentioned domain divergences either by using distribution matching methods [5], [6], [7], [8] or by transforming samples from one domain to another through generative models [9

Weak Adaptation Learning: Addressing Cross-domain Data Insufficiency with Weak Annotator

“…Previous Domain Adaptation (DA) approaches often design a number of losses to measure the discrepancy between source and target domains (Saito et al, 2018;Long et al, 2016;Tzeng et al, 2014;Zellinger et al, 2017;Roy et al, 2019). Then, adversarial training is applied in DA (Mathieu et al, 2016;Huang et al, 2018;Shermin et al, 2020;Tang & Jia, 2020). In the problem settings of these works, the supervision knowledge reflected in the source labels is accurate and reliable, which is not the case in our problem (where weak annotator generates inaccurate labels).…”

Decompose to Adapt: Cross-domain Object Detection via Feature Disentanglement

“…Unsupervised domain adaptation (UDA) aims at bridging the gap between an annotated source domain and an unannotated target domain. Typically, UDA methods transfer the knowledge in four ways: 1) directly minimizing the statistical distribution distance between two domains [22], [23]; 2) inducing the domain-invariant feature generation [9], [10], [24]; 3) learning from the synthesized images [25]- [29], and 4) selftraining via pseudo labels [30], [31]. By alleviating the crossdomain discrepancy at the feature and appearance levels, UDA methods have achieved outstanding performance in crossdomain classification [9], [10], [27], segmentation [32]- [34], and detection [11], [13], [14], [30], [35], [36].…”

Information Fusion Based on Multi-View and Multi-Feature Learning

“…They proposed a basic framework that leverages labeled base data and a saliency model to discover novel classes in unlabeled images through clustering. Shermin et al [24] introduced multiple classifiers within an adversarial framework to enhance domain adaptation in scenarios where target domain classes are incompletely known. These works contribute to the field of NCD by exploring different aspects such as image classification, 3D semantic segmentation, contrastive learning, and leveraging prior knowledge.…”