DINOv2<T>

Models & Types · AiDotNet.VisionLanguage.Encoders

DINOv2 self-supervised vision encoder producing universal visual features without labels.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → DINOv2<T>

Implements: IVisualEncoder<T>

For Beginners

DINOv2 from Meta learns powerful visual features without any labeled data using self-supervised training on 142 million curated images. Its features are so general that a simple linear classifier on top achieves strong results for image classification, segmentation, and depth estimation — making it a universal vision backbone. Default values follow the original paper settings.

How It Works

DINOv2 (Oquab et al., 2024) trains ViT with iBOT masked image modeling + DINO self-distillation on LVD-142M curated images. Register tokens reduce artifact patterns in attention maps. The model produces features usable for classification, segmentation, and depth estimation via linear probing.

References:

  • Paper: “DINOv2: Learning Robust Visual Features without Supervision” (Oquab et al., 2024)

Example

// Create a DINOv2 model for self-supervised visual feature extraction
// producing universal features for classification, segmentation, and depth
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
            
// ONNX inference mode with pre-trained model
var model = new DINOv2<double>(architecture, "dinov2.onnx");
            
// Training mode with native layers
var trainModel = new DINOv2<double>(architecture, new DINOv2Options());

Constructors

Constructor Summary
DINOv2(NeuralNetworkArchitecture<T> architecture, DINOv2Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
DINOv2(NeuralNetworkArchitecture<T> architecture, string modelPath, DINOv2Options options)

Properties

Property Summary
int EmbeddingDimension { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> EncodeImage(Tensor<T> image)