Pre-training YOLOv11 by Distilling DINOv3 скачать в хорошем качестве

Pre-training YOLOv11 by Distilling DINOv3

Meta has open-sourced DINOv3’s code and model weights (21M up to 7B parameters) under a generous license. You can load pre-trained DINOv3 backbones via PyTorch or Hugging Face and use for your tasks. DINOv3 (Distillation with No Labels v3) is self-supervised Vision Transformer (ViT), and knowledge distillation can be done without human annotations via LightlyTrain framework. Transfer the rich, general-purpose visual understanding of DINOv3 "teacher" model to a lightweight, efficient YOLOv11 "student" model 🚀. This pre-training step gives the YOLO model a significant head start, leading to better performance, faster convergence, and improved data efficiency when you later fine-tune it on your specific (and smaller) labeled dataset. Image shows mAP50:95(P) and FPS for INT8 and FP16 quantized YOLOv11 (large and nano) models after distillation and fine-tuning. Nano models achive close to 700FPS at mAP50:95(P) of 0.91 🤯🚀🚀 Check out the Python code if you want to pre-train your own YOLOv11 model with DINOv3's knowledge: Github Repo: https://github.com/farhanaugustine/DI...