Perturbing Dominant Feature Modes for Single Domain-Generalized Object Detection

This paper addresses the challenge of developing object detectors capable of generalizing to unseen domains using only a single source domain during training, a problem of paramount importance for real-world applications such as self-driving cars and unmanned aerial vehicles. We propose a method for single domain-generalized object detection (Single-DGOD) by simulating domain shifts in the feature space through perturbations of the dominant modes of low-level features. Our experimental results demonstrate that this approach provides a more effective way of diversifying the available source domain during training, out-performing existing methods by significant margins across several challenging domain shift scenarios. Compared to recent work, the proposed approach improves the mAP performance by 7.7%, 3.8%, and 6.5% for Clipart, Watercolor, and Comic respectively.