$k-\text{NN}$ embeded space conditioning for enhanced few-shot object detection

Few-shot learning has attracted significant scientific interest in the past decade due to its applicability to visual tasks with a natural long-tailed distribution such as object detection. This paper introduces a novel and flexible few-shot object detection approach which can be adapted effortlessly to any candidate-based object detection frame-work. In particular, our proposed $\boldsymbol{kFEW}$ component leverages a kNN retrieval technique over the regions of interest space to build both a class-distribution and a weighted aggregated embedding conditioned by the recovered neighbours. The obtained kNN feature representation is used to drive the training process without any additional trainable parameters as well as during inference time by steering the assumed confidence and the predicted box coordinates of the detection model. We perform extensive experiments and ablation studies on MS COCO and Pascal VOC proving its efficiency and state-of-the-art results (by a margin of 2.3 mAP points on MS COCO and by a margin of 2.5 mAP points on Pascal VOC) in the context of few-shot-object detection. Additionally, we demonstrate its versatility and ease-of-integration aspect by incorporating over competitive few-shot object detection methods and providing superior results.