RTDETROptions<T>

Options & Configuration · AiDotNet.Models.Options

Hyperparameters of RT-DETR (Zhao et al., “DETRs Beat YOLOs on Real-time Object Detection”, CVPR 2024). The defaults are the paper’s decoder and denoising values, which are the same for every size.

Inheritance: ModelOptions → ObjectDetectionOptions<T> → RTDETROptions<T>

For Beginners

The defaults reproduce the published model; change them only to experiment.

How It Works

Size selects the paper’s backbone and the hybrid encoder that goes with it:

  • Nano = R18 and Small = R34, with CSP expansion 0.5 and 3 or 4 decoder layers.
  • Medium = R50, with 6 decoder layers.
  • Large and XLarge = R101, whose encoder is 384 wide with a 2048 feed-forward.

The paper’s HGNetv2 backbones (RT-DETR-L/X) are not available.

Constructors

Constructor Summary
RTDETROptions() Creates the options with the paper’s values.

Properties

Property Summary
double BoxNoiseScale { get; set; } Box noise scale.
double LabelNoiseRatio { get; set; } Label noise ratio.
double PositionalTemperature { get; set; } Temperature of AIFI’s 2-D sin-cos positional encoding.
int DecoderFeedForwardDimension { get; set; } Decoder feed-forward width.
int DecoderHiddenDimension { get; set; } Decoder width (hidden_dim).
int DenoisingQueries { get; set; } Denoising query budget (num_denoising), split into positive/negative groups.
int NumHeads { get; set; } Attention heads in AIFI and the decoder.
int NumQueries { get; set; } Queries chosen by IoU-aware query selection.
int NumSamplingPoints { get; set; } Sampling points per head and level in the decoder’s deformable attention.