OWLv2<T>

Models & Types · AiDotNet.VisionLanguage.Grounding

OWLv2: self-training for scaling open-vocabulary detection.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → OWLv2<T>

Implements: IVisualGroundingModel<T>

For Beginners

OWLv2 is an improved version of OWL-ViT that achieves better open-vocabulary detection through self-training at web scale. Default values follow the original paper settings.

How It Works

OWLv2 (Minderer et al., 2023) scales open-vocabulary object detection through self-training on web-scale image-text data. Building on OWL-ViT, it uses a stronger ViT-L/14 backbone with 960px input resolution, pseudo-label self-training from an OWL-ViT teacher model, and objectness-calibrated scoring for better separation of true objects from background patches.

References:

  • Paper: “Scaling Open-Vocabulary Object Detection” (Google, 2023)

Example

// Create an OWLv2 model for scaled open-vocabulary detection
// with self-training and objectness-calibrated scoring
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 OWLv2<double>(architecture, "owlv2.onnx");
            
// Training mode with native layers
var trainModel = new OWLv2<double>(architecture, new OWLv2Options());

Constructors

Constructor Summary
OWLv2(NeuralNetworkArchitecture<T> architecture, OWLv2Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
OWLv2(NeuralNetworkArchitecture<T> architecture, string modelPath, OWLv2Options 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 OWLv2’s scaled open-vocabulary detection approach.