Domain Expansion and Boundary Growth for Open-Set Single-Source Domain Generalization

TL;DR

This article proposes a novel learning approach based on domain expansion and boundary growth to expand the scarce source samples and enlarge the boundaries across the known classes that indirectly broaden the boundary between the known and unknown classes.

Abstract

Open-set single-source domain generalization aims to use a single-source domain to learn a robust model that can be generalized to unknown target domains with both domain shifts and label shifts. The scarcity of the source domain and the unknown data distribution of the target domain pose a great challenge for domain-invariant feature learning and unknown class recognition. In this paper, we propose a novel learning approach based on domain expansion and boundary growth to expand the scarce source samples and enlarge the boundaries across the known classes that indirectly broaden the boundary between the known and unknown classes. Specifically, we achieve domain expansion by employing both background suppression and style augmentation on the source data to synthesize new samples. Then we force the model to distill consistent knowledge from the synthesized samples so that the model can learn domain-invariant information. Furthermore, we realize boundary growth across classes by using edge maps as an additional modality of samples when training multi-binary classifiers. In this way, it enlarges the boundary between the inliers and outliers, and consequently improves the unknown class recognition during open-set generalization. Extensive experiments show that our approach can achieve significant improvements and reach state-of-the-art performance on several cross-domain image classification datasets.

Figures (6)

• Figure 1: In the open-set single-source domain generalization (OS-SDG) setting, the capacity of our model in recognizing out-of-distribution known and open-set classes can be simultaneously enhanced when the known class distributions in the source feature space are adequately discriminative and well separated through domain expansion and boundary growth.
• Figure 2: The overall framework of our proposed DEBUG. Domain expansion is employed by applying background suppression and style augmentation and then computing the knowledge distillation loss $\mathcal{L}_{kd}$. Boundary growth is employed by extracting edge maps of the images and then using them as hard positive and negative samples to compute the edge-vs-all loss $\mathcal{L}_{eva}$.
• Figure 3: The t-SNE visualization of the encoded feature distribution of original images and edge maps. The original image samples are represented in deep colors, while the corresponding edge maps are depicted in light colors. The samples are drawn from the 'dog', 'giraffe', and 'guitar' categories of the three domains on PACS dataset.
• Figure 5: Comparison of Validation Accuracy (on the source domain validation set) and h-score (on the remaining unseen domains) between DEBUG and the DSU+OVA baseline.
• Figure 6: Per-class accuracy on unseen domains using different source domains in the PACS dataset.