Résumé
Recent neural networks for object detection achieve excellent performance when trained on large databases but still struggle to learn new objects with few examples, leading to the development of Few-Shot Object Detection (FSOD). State-of-the-art FSOD methods often use computationally heavy architectures like Faster R-CNN or pre-trained Vision Transformers (ViTs) such as DINOV2, limiting their usage for real-time inference in resource-constrained environments. We propose a novel approach that transfers rich features from a pre-trained ViT (DINOV2) to the lightweight YOLOv8 architecture via knowledge distillation. This bridges the gap between FSOD performance and efficiency, enabling YOLOv8 to better handle few-shot scenarios while retaining computational efficiency. Experiments on the MSCOCO benchmark adapted for FSOD show that our method enhances the performance of lightweight detectors, highlighting the benefits of combining ViT feature learning with efficient detectors for real-world FSOD.