meta segment anything model

After The Segment Anything Model, Meta Releases Another Open-Source Computer Vision Model Named DINOv2

It’s been thought that the AI battle is being played out mainly between Google and OpenAI, but Meta — in the last few weeks — has made some impressive new releases.

Meta has announced the release of DINOv2, which it says is a self-supervised Vision Transformer Model. DINOv2 are “a family of foundation models producing universal features suitable for image-level visual tasks (image classification, instance retrieval, video understanding) as well as pixel-level visual tasks (depth estimation, semantic segmentation),” Meta says.

“DINOv2 complements our other computer vision work, such as Segment Anything. While SAM is a promptable system focused on zero-shot generalization to diverse segmentation tasks, DINOv2 uses simple linear classifiers to achieve strong results across tasks beyond segmentation,” Meta says.

Meta said that it had already used DINOv2 to map entire forests, tree-by-tree, across areas the size of continents.

The model is open-source, and tinkerers will soon play around with it and report findings. But the release of DINOv2 follows the release of the extremely impressive Segment Anything Model (SAM), which Meta had released a few weeks ago. SAM was open-source with its weights available, and allows users to segment an image into its constituent parts. For instance, in an image of a dog eating from a bowl, the Segment Anything Model automatically draws “masks” around the dog, the bowl, and even individual features, like the dog’s collar. This has a variety of applications in image processing, such as with image editing, pattern matching, and image recognition.

These are all impressive steps, and Meta is even open sourcing many of these developments. For years, Google has been similarly open-sourcing its AI research — the transformers paper, which forms of the basis of many of today’s advances in AI, was an example of such open-sourced research. But while Meta had pivoted hard towards the Metaverse over the recent past, going as far as to change its name to Meta, recent developments show that the company is still carrying out some interesting new research, and could well be a dark horse in the AI race that’s likely to ensue in the coming years.

2 thoughts on “After The Segment Anything Model, Meta Releases Another Open-Source Computer Vision Model Named DINOv2

  1. Pingback: Apple Researchers Come Up With FaceLit To Create 3D Faces From 2D Pictures - MagicWand AI

  2. Pingback: Meta's New MMS Project Can Perform Speech-To-Text For 1,100 Languages - MagicWand AI

Leave a Reply

Your email address will not be published. Required fields are marked *