KDNet: Leveraging Vision-Language Knowledge Distillation for Few-Shot Object Detection

Few-shot object detection (FSOD) aims to detect new categories given only few instances for training. Recently emerged vision-language models (VLMs) have shown great performances in zero-shot and open-vocabulary object detection due to their strong ability to align object-level embedding with textual embedding of categories. However, few existing models distill VLMs' object-level knowledge in FSOD, which can help FSOD to learn novel semantic concepts to gain further improvement. Inspired by the recent knowledge distillation approaches with VLMs, we propose an end-to-end few-shot object detector with knowledge distillation from pre-trained VLMs, termed KDNet. A knowledge distillation branch is introduced alongside the object detector to distill knowledge from VLMs' visual encoder to the object detector. Also, we propose a pre-training mechanism with large-scale dataset to inject more semantic concepts to the detector to improve the performance on small datasets. The KDNet achieved state-of-the-art performance on both PASCAL VOC and MS COCO benchmarks over most of the shot settings and evaluation metrics.