This project is to create a pure C++ inference api for Segment Anything, MobileSAM and HQ-SAM, with no dependence on Python during runtime. The code repository contains a C++ library with a test program to facilitate easy integration of the interface into other projects.
Model loading takes approximately 10 or 1 seconds, and a single inference takes around 20ms, obtained using Intel Xeon W-2145 CPU (16 threads). During runtime, the interface may consume around 6GB or 1GB memory if running on CPU, and 16GB or 1GB if running on CUDA. The "or" here means values for "Segment Anything" or "MobileSAM".
Currently, this program has been thoroughly tested on Windows and may encounter issues when running on Linux (no pre-compiled Linux version is provided).
Video demo link
Image by Brenda from Pixabay and key recorded by KeyCastOW
Usage:
Download compressed file in the release page, unzip it, and run sam_cpp_test directly or in command line:
# Show help
./sam_cpp_test -h
# Example (change device, use CPU for preprocess and CUDA for sam)
# If you have multiple GPUs, you can use CUDA:1, CUDA:2, etc.
# All in cpu or all in cuda is also supported
./sam_cpp_test -pre_device="cpu" -sam_device="cuda:0"
# Example (default options)
./sam_cpp_test -pre_model="models/sam_preprocess.onnx" -sam_model="models/sam_vit_h_4b8939.onnx" -image="images/input.jpg"
# Example (use MobileSAM)
./sam_cpp_test -pre_model="models/mobile_sam_preprocess.onnx" -sam_model="models/mobile_sam.onnx"
# Example (use HQ-SAM)
./sam_cpp_test -pre_model="models/sam_hq_preprocess.onnx" -sam_model="models/sam_hq_vit_h.onnx"
# Example (change image)
./sam_cpp_test -image="images/input2.jpg"A simple example:
Sam::Parameter param("sam_preprocess.onnx", "sam_vit_h_4b8939.onnx", std::thread::hardware_concurrency());
param.providers[0].deviceType = 0; // cpu for preprocess
param.providers[1].deviceType = 1; // CUDA for sam
Sam sam(param);
// Use MobileSAM
Sam::Parameter param("mobile_sam_preprocess.onnx", "mobile_sam.onnx", std::thread::hardware_concurrency());
// Use HQ-SAM
Sam::Parameter param("sam_hq_preprocess.onnx", "sam_hq_vit_h.onnx", std::thread::hardware_concurrency());
auto inputSize = sam.getInputSize();
cv::Mat image = cv::imread("input.jpg", -1);
cv::resize(image, image, inputSize);
sam.loadImage(image); // Will require 6GB memory if using CPU, 16GB if using CUDA
// Using SAM with prompts (input: x, y)
cv::Mat mask = sam.getMask({200, 300});
cv::imwrite("output.png", mask);
// Using SAM with multiple prompts (input: points, nagativePoints)
cv::Mat mask = sam.getMask(points, nagativePoints); //Will require 1GB memory/graphics memory
cv::imwrite("output-multi.png", mask);
// Using SAM with box prompts (input: points, nagativePoints, box)
// The points and negativePoints can be empty (use {} as parameter)
cv::Rect box{444, 296, 171, 397};
cv::Mat mask = sam.getMask(points, nagativePoints, box);
cv::imwrite("output-box.png", mask);
// Automatically generating masks (input: number of points each side)
// Slow since running on CPU and the result is not as good as official demo
cv::Mat maskAuto = sam.autoSegment({10, 10});
cv::imwrite("output-auto.png", maskAuto);More details can be found in test.cpp and sam.h.
The "sam_vit_h_4b8939.onnx" and "mobile_sam.onnx" model can be exported using the official steps in here and here. The "sam_preprocess.onnx" and "mobile_sam_preprocess.onnx" models need to be exported using the export_pre_model script (see below).
Segment Anything involves several preprocessing steps, like this:
sam.to(device='cuda')
predictor = SamPredictor(sam)
predictor.set_image(image)
image_embedding = predictor.get_image_embedding().cpu().numpy()The export_pre_model script exports these operations as an ONNX model to enable execution independent of the Python environment. One limitation of this approach is that the exported model is dependent on a specific image size, so subsequent usage will require scaling images to that size. If you wish to modify the input image size (longest side not exceed 1024), the preprocessing model must be re-exported. Running the script requires installation of the Segment Anything and MobileSAM, and it requires approximately 23GB or 2GB of memory during execution for "Segment Anything" or "MobileSAM" respectively.
The export_pre_model script needs to be modified to switch between Segment-anything and MobileSAM:
output_names = ['output']
# Generate preprocessing model of Segment-anything in onnx format
# Use original segment-anything, Mobile-SAM or HQ-SAM model
# Each in a separate Python virtual environment
# Uncomment the following lines to generate preprocessing model of Segment-anything
# import segment_anything as SAM
# # Download Segment-anything model "sam_vit_h_4b8939.pth" from https://github.com/facebookresearch/segment-anything#model-checkpoints
# # and change the path below
# checkpoint = 'sam_vit_h_4b8939.pth'
# model_type = 'vit_h'
# output_path = 'models/sam_preprocess.onnx'
# quantize = True
# Mobile-SAM
# # Download Mobile-SAM model "mobile_sam.pt" from https://github.com/ChaoningZhang/MobileSAM/blob/master/weights/mobile_sam.pt
# import mobile_sam as SAM
# checkpoint = 'mobile_sam.pt'
# model_type = 'vit_t'
# output_path = 'models/mobile_sam_preprocess.onnx'
# quantize = False
# HQ-SAM
# # Download Mobile-SAM model "sam_hq_vit_h.pt" from https://github.com/SysCV/sam-hq#model-checkpoints
# # Installation: https://github.com/SysCV/sam-hq#quick-installation-via-pip
import segment_anything as SAM
checkpoint = 'sam_hq_vit_h.pth'
model_type = 'vit_h'
output_path = 'models/sam_hq_preprocess.onnx'
quantize = True
output_names = ['output', 'interm_embeddings']
# Target image size is 1024x720
image_size = (1024, 720)```bash
### Build
First, install the dependencies in [vcpkg](https://vcpkg.io):
#### Windows
```bash
./vcpkg install opencv:x64-windows gflags:x64-windows onnxruntime-gpu:x64-windowsThen, build the project with cmake.
mkdir build
cd build
cmake .. -DCMAKE_TOOLCHAIN_FILE=[vcpkg root]/scripts/buildsystems/vcpkg.cmakeDownload onnxruntime-linux-x64-1.14.1.tgz
./vcpkg install opencv:x64-linux gflags:x64-linuxbuild the project with cmake.
mkdir build
cd build
cmake .. -DCMAKE_TOOLCHAIN_FILE=[vcpkg root]/scripts/buildsystems/vcpkg.cmake -DONNXRUNTIME_ROOT_DIR=[onnxruntime-linux-x64-1.14.1 root]MIT