TextQueriedSegmentationOptions
Options & Configuration · AiDotNet.Models.Options
Options shared by segmentation models that classify pixels against text embeddings from a CLIP-style text encoder: the open-vocabulary models (CAT-Seg, SAN, SED, Mask-Adapter, Open-Vocabulary SAM) and the referring models (LISA, VideoLISA, OMG-LLaVA, PixelLM).
Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → TextQueriedSegmentationOptions
Known derived types: CATSegOptions, MaskAdapterOptions, OpenVocabSAMOptions, ReferringSegmentationOptions, SANOptions, SEDOptions
For Beginners
These models never learn a fixed list of classes. They turn each class name (or a sentence) into a vector with a text encoder, turn each pixel into a vector of the same size, and score a pixel against a name by how closely the two vectors point the same way.
Properties
| Property | Summary |
|---|---|
double Temperature { get; set; } |
Gets or sets the softmax temperature tau that divides each cosine similarity between a pixel embedding and a text embedding. |
string[] ClassNames { get; set; } |
Gets or sets the names of the classes the plain Predict/Train path scores against, one per output channel. |