We investigate the use of DINOv2 feature vectors for few-shot object detection, emphasizing methods for explainable and localized image analysis. By visualizing the similarity of local patch-based features with the class prototypes, we identify image regions that contribute to incorrect detections or are underrepresented, and reveal the importance of background information. Our evaluation shows that optimizing detection performance for individual classes does not necessarily improve overall performance, whereas incorporating explainability into the analysis can guide the development of methods that achieve measurable performance gains. These findings provide insight into the decision-making process of the used few-shot detector and demonstrate the value of visual explanations for improving model robustness and interpretability.