MiDaS<T>

Models & Types · AiDotNet.Video.Depth

MiDaS: Towards Robust Monocular Depth Estimation.

Inheritance: NeuralNetworkBase<T> → MiDaS<T>

For Beginners

MiDaS estimates depth from a single image (monocular depth estimation). Unlike stereo vision which uses two cameras, MiDaS uses deep learning to predict depth from visual cues like texture, perspective, and object sizes.

Key capabilities:

  • Single image to depth map conversion
  • Works on arbitrary images without camera parameters
  • Outputs relative depth (closer objects have higher values)
  • Robust across different scenes and domains

Example usage:

How It Works

Technical Details:

  • ViT-based encoder for feature extraction
  • Multi-scale fusion decoder
  • Scale and shift invariant loss for training
  • Trained on diverse mixed datasets for robustness

Reference: “Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer” https://arxiv.org/abs/1907.01341

Constructors

Constructor Summary
MiDaS() Creates a MiDaS model with default configuration.
MiDaS(NeuralNetworkArchitecture<T> architecture, MiDaSOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a MiDaS model using native layers for training and inference.
MiDaS(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, MiDaSOptions options) Creates a MiDaS model using a pretrained ONNX model for inference.

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
List<Tensor<T>> EstimateDepthForVideo(List<Tensor<T>> frames) Estimates depth for multiple video frames.
ModelMetadata<T> GetModelMetadata()
Tensor<T> EstimateDepth(Tensor<T> image) Estimates depth from an input image.
Tensor<T> NormalizeDepthMap(Tensor<T> depthMap) Normalizes depth map to 0-1 range for visualization.