reusex::vision::onnx::ONNXSam3 Class Reference

SAM3 segmentation model using ONNX Runtime for CPU/GPU inference. More...

#include <Sam3.hpp>

Inheritance diagram for reusex::vision::onnx::ONNXSam3:

Collaboration diagram for reusex::vision::onnx::ONNXSam3:

Public Member Functions
	ONNXSam3 (const std::filesystem::path &model_dir, bool use_cuda=false)
	Construct and load a SAM3 model from a directory.
std::vector< IDataset::Pair >	forward (const std::span< IDataset::Pair > &input) override
	Run SAM3 inference on a batch of images.
Public Member Functions inherited from reusex::vision::IModel
virtual	~IModel ()=default

Static Public Member Functions
static std::unique_ptr< ONNXSam3 >	create (const std::filesystem::path &model_dir, bool use_cuda=false)
	Factory method for creating an ONNXSam3 instance.
Static Public Member Functions inherited from reusex::vision::IModel
static std::unique_ptr< IModel >	create (const std::filesystem::path &model_path, bool use_gpu=false)

Detailed Description

SAM3 segmentation model using ONNX Runtime for CPU/GPU inference.

Loads three ONNX sub-models (vision encoder, text encoder, decoder) from a directory and implements the IModel interface for use with the backend-agnostic annotation pipeline.

Optionally uses CUDA execution provider when available, with automatic fallback to CPU.

Definition at line 30 of file Sam3.hpp.

Constructor & Destructor Documentation

ONNXSam3()

reusex::vision::onnx::ONNXSam3::ONNXSam3 ( const std::filesystem::path & model_dir, bool use_cuda = false )

explicit

Construct and load a SAM3 model from a directory.

Parameters

model_dir	Directory containing vision-encoder.onnx, text-encoder.onnx, decoder.onnx, and tokenizer.json.
use_cuda	Whether to attempt CUDA execution provider.

Member Function Documentation

create()

std::unique_ptr< ONNXSam3 > reusex::vision::onnx::ONNXSam3::create ( const std::filesystem::path & model_dir, bool use_cuda = false )

static

Factory method for creating an ONNXSam3 instance.

Parameters

model_dir	Directory with ONNX model files.
use_cuda	Whether to use CUDA (defaults to false).

Returns

Unique pointer to the created model.

forward()

std::vector< IDataset::Pair > reusex::vision::onnx::ONNXSam3::forward ( const std::span< IDataset::Pair > & input )

overridevirtual

Run SAM3 inference on a batch of images.

Each input Pair must contain an ONNXSam3Data with a raw image and text prompts. Output Pairs contain the same data type with the image field set to the label image (CV_32S, -1 for background).

Parameters

input

Span of (data, index) pairs from the dataset.

Returns

Vector of (data, index) pairs with label images.

Implements reusex::vision::IModel.

---

The documentation for this class was generated from the following file:

• libs/reusex/include/vision/onnx/Sam3.hpp