OWLViT<T>

Models & Types · AiDotNet.VisionLanguage.Grounding

OWL-ViT: open-vocabulary object detection via ViT + CLIP alignment.

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

Implements: IVisualGroundingModel<T>

For Beginners

OWL-ViT is a vision-language model that detects objects using text descriptions by leveraging CLIP’s vision-language alignment at the patch level. Default values follow the original paper settings.

How It Works

OWL-ViT (Minderer et al., 2022) repurposes a CLIP ViT image encoder for open-vocabulary detection by treating each patch token as a candidate object. Per-patch MLP heads predict bounding boxes, while class scores come from cosine similarity between patch tokens and CLIP text embeddings of category names. This is a query-free approach where every ViT patch serves as a candidate detection, eliminating the need for learned object queries.

References:

  • Paper: “Simple Open-Vocabulary Object Detection with Vision Transformers” (Google, 2022)

Example

// Create an OWL-ViT model for open-vocabulary object detection
// using CLIP-aligned patch-level detection without learned queries
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 OWLViT<double>(architecture, "owlvit.onnx");
            
// Training mode with native layers
var trainModel = new OWLViT<double>(architecture, new OWLViTOptions());

Constructors

Constructor Summary
OWLViT(NeuralNetworkArchitecture<T> architecture, OWLViTOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
OWLViT(NeuralNetworkArchitecture<T> architecture, string modelPath, OWLViTOptions 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 OWL-ViT’s CLIP-aligned patch-level detection approach.