Co-Teaching for Unsupervised Domain Adaptation and Expansion

TL;DR

This work addresses the limitations of Unsupervised Domain Adaptation by proposing Unsupervised Domain Expansion, which preserves source-domain performance while adapting to a target domain. It introduces Co-Teaching (CT), a two-stage framework combining knowledge-distillation based co-teaching (kdCT) and mixup based co-teaching (miCT) to handle cross-domain ambiguity via biased teacher guidance and cross-domain data augmentation. CT leverages two domain-specific teachers to guide a student, with stochastic weighting and mixup to improve generalization, and demonstrates strong performance gains on multi-class image classification (Office-Home) and driving-scene segmentation (Cityscapes/ACDC) compared to KDDE and UDA baselines. The results suggest cross-domain ambiguity is a key lever for improving domain expansion, offering a practical, backbone-agnostic baseline for real-world deployment where single-domain specialization is insufficient.

Abstract

Unsupervised Domain Adaptation (UDA) essentially trades a model's performance on a source domain for improving its performance on a target domain. To overcome this, Unsupervised Domain Expansion (UDE) has been introduced, which adapts the model to the target domain while preserving its performance in the source domain. In both UDA and UDE, a model tailored to a given domain is assumed to well handle samples from the given domain. We question the assumption by reporting the existence of cross-domain visual ambiguity: Due to the unclear boundary between the two domains, samples from one domain can be visually close to the other domain. Such sorts of samples are typically in the minority in their host domain, so they tend to be overlooked by the domain-specific model, but can be better handled by a model from the other domain. We exploit this finding by proposing Co-Teaching (CT), which is instantiated with knowledge distillation based CT (kdCT) plus mixup based CT (miCT). Specifically, kdCT leverages a dual-teacher architecture to enhance the student network's ability to handle cross-domain ambiguity. Meanwhile, miCT further enhances the generalization ability of the student. Extensive experiments on image classification and driving-scene segmentation show the viability of CT for UDE.

Figures (4)

• Figure 1: Cross-domain visual ambiguity. Samples from a target domain (Clipart) can be visually realistic as samples from a source domain (Art), and vice versa.
• Figure 2: Proposed CT method. Given labeled data $\{(x_s, y_s)\}$ from a source domain and unlabeled data $\{x_t\}$ from a target domain, CT obtains a domain-expanded network $\mathcal{N}_u$ by two-stage training. In the first stage, two domain-specific teacher networks $\mathcal{N}_s$ and $\mathcal{N}_t$ are obtained, where $\mathcal{N}_s$ for the source domain is trained on $\{(x_s,y_s)\}$ by standard supervised learning, whilst $\mathcal{N}_t$ for the target domain is trained on $\{(x_s,y_s)\}$ and $\{x_t\}$ by an existing UDA method. In the second stage, $\mathcal{N}_s$ and $\mathcal{N}_t$ co-teach $\mathcal{N}_u$ via knowledge distillation based CT (kdCT) that minimizes $L_{kdct}(\{x_s\}) + L_{kdct}(\{x_t\})$, and mixup based CT (miCT), minimizing $L_{mict}(\{x_m\})$ for mixup instances $\{x_m\}$. Once trained, only $\mathcal{N}_u$ is needed for inference.
• Figure 3: Grad-CAM gradcam visualization. The top three rows are from a source domain (Art), while the bottom three rows are from a target domain (Clipart). Texts under heatmaps are predicted labels with scores.
• Figure 4: Qualitative results of driving scene segmentation. The first row is from $D_s$ (normal condition), while the other rows are from $D_t$ (adverse conditions in the nighttime, fog, snow and rain). Important difference between the results is marked out by white bounding boxes. Best viewed digitally.