Beyond the Known: Enhancing Open Set Domain Adaptation with Unknown Exploration

TL;DR

This work tackles Open Set Domain Adaptation (OSDA) by leveraging high-confidence unknown target samples to tighten the decision boundaries of an OVANet-based model. It introduces three strategies—Original, Augmentation, and Generation (with GAN-generated negatives and entropy/agreement losses)—to incorporate negative supervision via a new loss $\mathcal{L}_{ova}^{neg}$. A discriminator-augmented Generation++ approach with interleaved training further refines synthetic negatives, guided by $\mathcal{L}_{ent}^{gen}$ and $\mathcal{L}_{agree}^{gen}$. Experiments on Office-31, Office-Home, and VisDA show competitive $H$-scores while boosting unknown-category accuracy, and the method is presented as widely applicable to OSDA baselines beyond OVANet. The findings highlight a practical path to improve unknown rejection in OSDA through targeted exploitation of negative information.

Abstract

Convolutional neural networks (CNNs) can learn directly from raw data, resulting in exceptional performance across various research areas. However, factors present in non-controllable environments such as unlabeled datasets with varying levels of domain and category shift can reduce model accuracy. The Open Set Domain Adaptation (OSDA) is a challenging problem that arises when both of these issues occur together. Existing OSDA approaches in literature only align known classes or use supervised training to learn unknown classes as a single new category. In this work, we introduce a new approach to improve OSDA techniques by extracting a set of high-confidence unknown instances and using it as a hard constraint to tighten the classification boundaries. Specifically, we use a new loss constraint that is evaluated in three different ways: (1) using pristine negative instances directly; (2) using data augmentation techniques to create randomly transformed negatives; and (3) with generated synthetic negatives containing adversarial features. We analyze different strategies to improve the discriminator and the training of the Generative Adversarial Network (GAN) used to generate synthetic negatives. We conducted extensive experiments and analysis on OVANet using three widely-used public benchmarks, the Office-31, Office-Home, and VisDA datasets. We were able to achieve similar H-score to other state-of-the-art methods, while increasing the accuracy on unknown categories.

Figures (7)

• Figure 1: Top-level overview of our approach. Initially, a base method, in this particular case OVANet, is trained for half of the expected iterations. Then, at the breakpoint, a sample extraction routine is executed to select the set of the most confident negative samples $\mathbf{\overline{X}}$ from the target domain. Next, the base method leverages the negative samples $\mathbf{\overline{X}}$ as a new constraint $\mathcal{L}_{ova}^{neg}$ for the last half of the iterations to tighten the classification boundaries.
• Figure 2: OVANet inference procedure diagram. $G$ refers to the feature extraction entity, in our particular case the ResNet50 model; $C$ consists of the CS head, and $O$ refers to the one-vs-all set of $|L_s|$ binary classifiers of the OS head.
• Figure 3: Description of all three strategies. (a) refers to the original approach, which feeds the instances $\overline{x}_i \in \overline{\mathbf{X}}$ directly to OVANet. (b) depicts the augmentation approach, which applies Random Affine and Gaussian Blur transformations to the instances $\overline{x}_i \in \overline{\mathbf{X}}$ before feeding them to OVANet. (c) shows the generation approach, where a DCGAN model composed of a generator (Gen) and a discriminator (Disc) is used. Gen is trained to fool both the Disc and OVANet models by outputting synthetic negative instances aligned to the known classes in $L_s$, while Disc distinguishes between real and fake inputs. The dashed arrow indicates the backpropagation step, underscoring the dependence between the Gen entity and the loss error from the Disc and OVANet models.
• Figure 4: Distribution of the obtained probabilities $1-p_o(\hat{y}_i^{k} | \mathbf{x}_i)$ for unknown instances on the target training set of the Office-31 dataset. Each of the boxplots refers to a specific task and the red line shows the adopted threshold of 0.9.
• Figure 5: Ilustrations of the OVANet's feature extraction as the backbone of the GAN discriminator.