Fusing Foundation Models with Fast Object Detectors

YoDINO is an experimental computer vision project exploring the integration of massive self-supervised Vision Transformers (specifically frozen DINOv3 backbones) with the high-speed, light-weight detection heads of the YOLO family.

By leveraging a neck network design, such as Simple Feature Pyramid (SFP), to translate ViT attention representations into multi-scale feature maps, YoDINO achieves high detection precision in small data environments. This framework allows developers to capitalize on the rich feature representation of foundation vision models while maintaining real-time object detection capabilities.