LEOVL<T>

Models & Types · AiDotNet.VisionLanguage.ThreeD

LEO-VL: efficient 3D scene representation from multi-view RGB-D.

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

Implements: IThreeDVisionLanguageModel<T>

For Beginners

LEO-VL is a vision-language model for 3D scene understanding using multi-view RGB-D camera inputs. Default values follow the original paper settings.

How It Works

LEO-VL (2025) provides efficient 3D scene understanding from multi-view RGB-D inputs. It constructs 3D scene representations from multiple depth-camera views without requiring explicit 3D reconstruction, using a view aggregation module to fuse multi-view features into a coherent spatial representation for language-guided scene understanding and navigation.

References:

  • Paper: “LEO-VL: Efficient 3D Scene Understanding via Multi-View RGB-D (Various, 2025)”

Example

// Create a LEO-VL model for 3D scene understanding
// with efficient multi-view RGB-D representation learning
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 LEOVL<double>(architecture, "leovl.onnx");
            
// Training mode with native layers
var trainModel = new LEOVL<double>(architecture, new LEOVLOptions());

Constructors

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

Properties

Property Summary
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
int MaxPoints { get; }
int PointChannels { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFrom3D(Tensor<T> pointCloud, string prompt) Processes 3D point cloud using LEO-VL’s embodied spatial reasoning approach.
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates from 2D image using LEO-VL’s embodied multi-modal approach.