RobustNet: Improving Domain Generalization in Urban-Scene Segmentation via Instance Selective Whitening

Enhancing the generalization capability of deep neural networks to unseen\ndomains is crucial for safety-critical applications in the real world such as\nautonomous driving. To address this issue, this paper proposes a novel instance\nselective whitening loss to improve the robustness of the segmentation networks\nfor unseen domains. Our approach disentangles the domain-specific style and\ndomain-invariant content encoded in higher-order statistics (i.e., feature\ncovariance) of the feature representations and selectively removes only the\nstyle information causing domain shift. As shown in Fig. 1, our method provides\nreasonable predictions for (a) low-illuminated, (b) rainy, and (c) unseen\nstructures. These types of images are not included in the training dataset,\nwhere the baseline shows a significant performance drop, contrary to ours.\nBeing simple yet effective, our approach improves the robustness of various\nbackbone networks without additional computational cost. We conduct extensive\nexperiments in urban-scene segmentation and show the superiority of our\napproach to existing work. Our code is available at\nhttps://github.com/shachoi/RobustNet.\n