Domain Generalization through Attenuation of Domain-Specific Information

TL;DR

This work tackles domain generalization for automotive semantic segmentation by introducing Domain Independence (DI) to quantify where domain-specific information resides and Attenuation of Domain-Specific Information (ADSI) to suppress such cues. DI analyzes feature representations from a frozen encoder and a frequency-space decomposition to identify domain dependence, while ADSI applies a Butterworth-based low-frequency attenuation to both amplitude and phase components, with a color-preserving scalar in [0,1], enabling domain-robust learning from a single domain. Empirically, ADSI outperforms strong baselines like Rein across GTA5-to-Real and Cityscapes-to-ACDC transfers, and ablations show joint attenuation of amplitude and phase with Butterworth masks yields the best gains, though challenges remain under rain/night conditions. The proposed framework provides a principled, frequency-aware approach to domain generalization that leverages single-domain data while maintaining essential color information, with practical impact for robust perception in diverse driving environments.

Abstract

In this paper, we propose a new evaluation metric called Domain Independence (DI) and Attenuation of Domain-Specific Information (ADSI) which is specifically designed for domain-generalized semantic segmentation in automotive images. DI measures the presence of domain-specific information: a lower DI value indicates strong domain dependence, while a higher DI value suggests greater domain independence. This makes it roughly where domain-specific information exists and up to which frequency range it is present. As a result, it becomes possible to effectively suppress only the regions in the image that contain domain-specific information, enabling feature extraction independent of the domain. ADSI uses a Butterworth filter to remove the low-frequency components of images that contain inherent domain-specific information such as sensor characteristics and lighting conditions. However, since low-frequency components also contain important information such as color, we should not remove them completely. Thus, a scalar value (ranging from 0 to 1) is multiplied by the low-frequency components to retain essential information. This helps the model learn more domain-independent features. In experiments, GTA5 (synthetic dataset) was used as training images, and a real-world dataset was used for evaluation, and the proposed method outperformed conventional approaches. Similarly, in experiments that the Cityscapes (real-world dataset) was used for training and various environment datasets such as rain and nighttime were used for evaluation, the proposed method demonstrated its robustness under nighttime conditions.

Figures (8)

• Figure 1: The areas improved by the proposed method are highlighted. The red box indicates a bicycle hidden in the shadow, making it quite difficult to see. As a result, the conventional method, Rein, fails to recognize the bicycle. However, by applying the proposed method (Ours), the bicycle hidden in the shadow becomes recognizable.
• Figure 2: Definition of Domain-Specific and Domain-Independent. If the data has specific features to a particular domain, Domain-Independent metric will be low. On the other hand, if it does not contain domain-specific information, Domain-Independent metric will be high.
• Figure 3: Overview of Domain Independence metric. The image is passed through a frozen Image Encoder to extract its features. The reason the image encoder is frozen is that it ensures consistent feature extraction, allowing roughly evaluation of differences between image domains. After that, by comparing the distances between features, the degree to which the image depends on a specific domain can be quantified.
• Figure 4: DI image diagram and calculation method. For a certain feature, if the closest feature values are in the same dataset, we count the black lines. On the other hand, if they are in different datasets, count the red lines.
• Figure 5: Overview of low-frequency component removal. By removing the frequency components inside the box, sensor characteristics and lighting conditions present in the low-frequency components can be eliminated. Since filters like the Butterworth filter gradually reduce the frequency components, it is not clear where the domain-specific information is contained. Therefore, here we use a box filter to clearly eliminate the frequency components in specific areas. Additionally, by adjusting $\beta$, the size of the box can be modified.