ReferringSegmentationOptions

Options & Configuration · AiDotNet.Models.Options

Options shared by the referring segmentation models (LISA, VideoLISA, OMG-LLaVA, PixelLM), which turn a referring expression into one text embedding and decode a binary mask from it.

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → TextQueriedSegmentationOptions → ReferringSegmentationOptions

Known derived types: LISAOptions, OMGLLaVAOptions, PixelLMOptions, VideoLISAOptions

For Beginners

A referring model is given a sentence such as “the dog on the left” and returns the pixels it describes. All four models share the same last step: the sentence becomes one text embedding, each pixel’s embedding is compared with it, and the resulting score decides whether the pixel is in the mask. These options control that shared step and the text encoding (inherited from TextQueriedSegmentationOptions). Network sizes, such as encoder widths and depths, are configured on each model’s own options class.

How It Works

Reference: Lai et al., “LISA: Reasoning Segmentation via Large Language Model”, CVPR 2024.

Properties

Property Summary
double MaskThreshold { get; set; } Gets or sets the probability a pixel’s sigmoid mask score must exceed to be in the mask.