GroundingDINOOptions

Options & Configuration · AiDotNet.VisionLanguage.Grounding

Configuration options for Grounding DINO (Liu et al. 2023): open-set detection combining DINO with grounded pre-training.

Inheritance: ModelOptions → GenerativeVLMOptions → GroundingVLMOptions → GroundingDINOOptions

For Beginners

These options configure the GroundingDINO model. Default values follow the released model.

How It Works

The defaults are Grounding DINO-T as released (the official GroundingDINO_SwinT_OGC configuration): a Swin-T image backbone, a BERT-base-uncased text backbone, a 256-wide feature enhancer and decoder of six layers each with eight heads, four feature levels, four sampling points, 900 queries and a 256-token text head. The shared vision-language options map as follows: VisionDim is the Swin embedding width, DecoderDim the model width, NumVisionLayers the feature-enhancer layers, NumHeads the encoder and decoder heads, MaxSequenceLength the scored text tokens, VocabSize the BERT vocabulary and ConfidenceThreshold the box threshold.

Constructors

Constructor Summary
GroundingDINOOptions() Initializes a new instance with default values.
GroundingDINOOptions(GroundingDINOOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
double FusionDropPathRate { get; set; } Gets or sets the stochastic depth rate of the vision-text fusion layers (0.1).
double SwinDropPathRate { get; set; } Gets or sets the Swin stochastic depth rate (0.2, the official backbone’s).
double TextThreshold { get; set; } Gets or sets the token probability above which a token joins a detection’s phrase (0.25, the official demo’s text threshold).
int FeedForwardDim { get; set; } Gets or sets the feed-forward width of the feature enhancer and decoder layers (2048).
int NumFeatureLevels { get; set; } Gets or sets the feature levels: the backbone’s three plus one stride-2 convolution of the last (4).
int NumQueryPositions { get; set; } Gets or sets the number of object queries the decoder refines (900).
int NumSamplingPoints { get; set; } Gets or sets the deformable-attention sampling points per head and level (4).
int SwinWindowSize { get; set; } Gets or sets the Swin attention window (7).
int TextHeads { get; set; } Gets or sets the BERT attention heads (12).
int TextHiddenSize { get; set; } Gets or sets the BERT width (768).
int TextIntermediateSize { get; set; } Gets or sets the BERT feed-forward width (3072).
int TextLayers { get; set; } Gets or sets the BERT layers (12).
int[] SwinDepths { get; set; } Gets or sets the Swin blocks per stage (Swin-T: 2, 2, 6, 2).
int[] SwinHeads { get; set; } Gets or sets the Swin attention heads per stage (Swin-T: 3, 6, 12, 24).