FerretV2<T>

Models & Types · AiDotNet.VisionLanguage.Grounding

Ferret-v2: improved referring and grounding with enhanced spatial understanding.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → FerretV2<T>

Implements: IVisualGroundingModel<T>

For Beginners

Ferret-v2 is an improved vision-language model for referring and grounding that handles images at multiple resolutions for better spatial understanding. Default values follow the original paper settings.

How It Works

Ferret-v2 (Zhang et al., 2024) improves upon Ferret with multi-granularity visual encoding that processes images at multiple resolutions, enhanced spatial understanding through DINOv2 features, and a flexible any-resolution scaling approach for handling diverse aspect ratios and input sizes. The multi-granularity design captures both fine-grained details and global context.

References:

  • Paper: “Ferret-v2: An Improved Baseline for Referring and Grounding with Multi-Granularity Visual Encoding” (Apple, 2024)

Example

// Create a Ferret-v2 model for improved referring and grounding
// with multi-granularity visual encoding and enhanced spatial understanding
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 FerretV2<double>(architecture, "ferretv2.onnx");
            
// Training mode with native layers
var trainModel = new FerretV2<double>(architecture, new FerretV2Options());

Constructors

Constructor Summary
FerretV2(NeuralNetworkArchitecture<T> architecture, FerretV2Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
FerretV2(NeuralNetworkArchitecture<T> architecture, string modelPath, FerretV2Options options)

Properties

Property Summary
int EmbeddingDimension { get; }
int MaxDetections { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> DetectObjects(Tensor<T> image, IReadOnlyList<string> categories)
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GroundText(Tensor<T> image, string textQuery) Grounds text using Ferret-v2’s any-resolution high-res grounding approach.