A high-performance C++ implementation of the Bilateral Reference Network (BiRefNet) leveraging TensorRT and CUDA, optimized for real-time, high-resolution dichotomous image segmentation.

BiRefNet C++ TENSORRT

BiRefNet C++ TENSORRT is designed to efficiently run bilateral reference segmentation tasks on the GPU. By harnessing TensorRT’s optimizations and CUDA kernels, it aims to deliver state-of-the-art performance with minimal latency.

BiRefNet C++ TENSORRT BiRefNet C++ TENSORRT

• TensorRT Acceleration: Speed up inference for segmentation tasks using serialized TRT engines.
• CUDA Integration: Comprehensive GPU-based preprocessing, postprocessing, and memory handling.
• High-Resolution Support: Out-of-the-box ability to process high-resolution images (e.g., 1024x1024).
• Easy Integration: C++17 codebase for easy deployment into existing pipelines.

• Enhanced Bilateral Reference: Improves dichotomous segmentation outputs by leveraging dual reference guidance.
• Improved Memory Footprint: Optimized GPU allocation for large-batch or high-resolution workloads.
• Configurable Precision: Support for FP16 or FP32 modes (requires GPU with half-precision support for FP16).
• Flexible I/O: Easily integrate your own data loaders or pipeline steps thanks to modular design.

BiRefNet/ 
├── include 
│ └── birefnet.h # Main BiRefNet class definition
├── src 
│ └── birefnet.cpp # Implementation of the BiRefNet class
├── CMakeLists.txt # CMake configuration 
└── main.cpp #

• include/birefnet.h
  Header file defining the BiRefNet class, which manages TensorRT engine creation, execution, and memory buffers.

• src/birefnet.cpp
  Source implementation for loading serialized engines, running inference, and handling output postprocessing.

• CMakeLists.txt
  Configuration for building the project using CMake. Adjust paths to TensorRT, CUDA, and OpenCV as needed.

• main.cpp
  A minimal example demonstrating how to load the model, run inference on images or videos, and save the results.

1. Clone the Repository

    git clone https://github.com/hamdiboukamcha/BiRefNet-Cpp-TensorRT.git
    cd BiRefNet-Cpp-TensorRT
    mkdir build && cd build
    cmake ..
    make -j$(nproc)
   

CUDA Required for GPU acceleration and kernel launches (e.g., CUDA 11.x or later).

TensorRT High-performance deep learning inference library (v8.x or later recommended).

OpenCV Needed for image loading, preprocessing, and basic visualization.

C++17 This project uses modern C++ features. Ensure your compiler supports C++17 or above.

Main Components BiRefNet Class

Initializes a TensorRT engine from a given engine/model path. Handles preprocessing (image resizing, mean/std normalization, etc.). Runs inference and postprocesses outputs into segmentation maps. Manages CUDA resources and streams. Logger Class (in main.cpp)

Implements TensorRT’s ILogger interface for custom logging. Notable Functions BiRefNet::BiRefNet(...)

Constructor that loads a .trt (serialized TensorRT) engine into memory. BiRefNet::predict(cv::Mat& image)

Main function to run inference: takes an OpenCV cv::Mat as input, returns the segmented result as cv::Mat. BiRefNet::preprocess(...)

Converts an image into normalized floats (mean subtraction, division by std, etc.). BiRefNet::postprocess(...)

Reshapes the raw output into meaningful image data, typically an 8-bit or 32-bit matrix for segmentation.

Prepare Your Engine
Convert your model to ONNX, then build a TensorRT engine (e.g., using trtexec or a custom builder).
trtexec --onnx=birefnet_model.onnx --saveEngine=BiRefNet-tiny.engine --fp16

For advanced inquiries, feel free to contact me on LinkedIn: LinkedIn

If you use BiRefNet C++ TENSORRT in your academic work or research, please cite:

@misc{Boukamcha2025BiRefNet,
		author = {Hamdi Boukamcha},
		title = {BiRefNet C++ TENSORRT},
		year = {2025},
		publisher = {GitHub},
		howpublished = {\url{https://github.com/hamdiboukamcha/BiRefNet-Cpp-TensorRT}}
}