CDFormer: Cross-Domain Few-Shot Object Detection Transformer Against Feature Confusion

The object-background confusion refers to the confusion between expected objects and background. As illustrated in Figure 1(a), in underwater scenes, the boundaries between the target object and the background are often ambiguous, leading to missed detections. The object-object confusion refers to the confusion between different classes of objects. As illustrated in Figure 1(b), the similarity between different classes results in false detections. In the field of CD-FSOD, CD-ViTO [8] represents the state-of-the-art work, which devises various fine-tuning modules and achieves significant performance improvements. To address object-background confusion, CD-ViTO re-weights manually selected background features and combines them with object features in a weighted sum. However, manually designed features lack adaptability when the target domain distribution differs [4], [24]. To address object-object confusion, CD-ViTO [8] enhances class distinction by directly adjusting the support class features.