Posted by admin

In April 2023, Meta AI released DINOv2, a foundational computer vision model pretrained using self-supervised learning on a large curated dataset, which can be used without finetuning! In this video we'll examine interesting topics from this exciting release. If you enjoy this video then please subscribe to the channel and hit the like button to support the creation of more similar educational content. We'll start by explaining what is a foundational model and why DINOv2 can be counted as such. Next we'll see how you may use DINOv2 in your code. DINOv2 was released with few versions of different model sizes. We'll explain how Meta AI have created the smaller models using model distillation. We then review parts of the process of creating the large curated dataset that was used to train DINOv2. We'll finish by talking about how using self-supervised learning helped DINOv2 reach amazing pixel level understanding of the image, comparing to text guided images that are commonly used. GitHub repo - https://github.com/facebookresearch/dinov2 arxiv paper - https://arxiv.org/abs/2304.07193 ----------------------------------------------------------------------------------------------------------------------- https://www.patreon.com/CSBoard https://ko-fi.com/csboard ----------------------------------------------------------------------------------------------------------------------- Chapters: 0:00 Introduction 0:52 Foundational Model 2:45 Using DINOv2 3:10 Model Distillation 4:41 SSL with Curated Data 6:21 Pixel level learning