Few-Shot Object Detection has received strong interest recently, especially as collecting annotated training data for new and varied problems becomes time-consuming and expensive. In Few-Shot Object Detection, after training on ample “base” class data, the detector is finetuned on few examples of “novel” classes such that the inference performance is maximized over all classes in a test set. Most previous works aim to learn novel classes independently, without regard to interactions with semantically similar classes. We take a different view, integrating contrastive learning with synthesized positive samples from similar but distinct classes. We generate new positive samples via a convex combination between a novel class feature and its closest classes’ features; then, we encourage the novel class feature to be close to the generated features using contrastive learning. Our method achieves state-of-the-art results on the standard PASCAL-VOC and MS-COCO benchmarks, and does so with minimally extra learnable parameters during training and no extra parameters during inference.