Foundationpose-CPP

Windows

First, thanks to the original author. Since my deployment environment is Windows, I modified the original repo to run on Windows 10/11. Inference environment: CUDA 11.8 + CUDNN 8.9 + TensorRT 8.6.1.6

1. Basic Dependencies

For convenience, use vcpkg. Refer to the official documentation for usage:

git clone https://github.com/microsoft/vcpkg.git
cd vcpkg
./bootstrap-vcpkg.bat
./vcpkg.exe install assimp opencv4 eigen3 gtest glog 

2. CV-CUDA -> ppl.cv

CV-CUDA currently (2025-07-30) does not support Windows. The main modification in this branch is using ppl.cv instead. Compile ppl.cv:

git clone https://github.com/openppl/ppl.cv.git
cd ppl.cv
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../install ^
-DPYBIND11_NOPYTHON=ON -DPPLCV_USE_CUDA=ON -DPPLCV_USE_MSVC_STATIC_RUNTIME=OFF
cmake --build . --config Release --target ALL_BUILD
cmake --build . --config Release --target install

3. Compile This Repository

git clone https://github.com/wawatt/foundationpose_cpp_windows.git
cd foundationpose_cpp
git submodule init
git submodule update

Remove MSVC-unsupported flags -Wextra -Wdeprecated in: easy_deploy_tool\deploy_core and easy_deploy_tool\inference_core\trt_core

cd foundationpose_cpp
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=D:/build/vcpkg/scripts/buildsystems/vcpkg.cmake ^
-DENABLE_TENSORRT=ON -DTRT_DIR=E:/ENVs/TensorRT-8.6.1.6 ^
-Dpplcv_DIR="E:/ENVs/ppl.cv/install/lib/cmake/ppl" ^
-DCMAKE_CUDA_ARCHITECTURES="86;89"
cmake --build . --config Release --target ALL_BUILD

4. Convert ONNX to TensorRT Models

Script

5. Run

cd foundationpose_cpp_windows
"build/bin/Release/simple_tests.exe" --gtest_filter=foundationpose_test.test

The following text is unmodified...

About this project

This project is adapted from nvidia-issac-pose-estimation, with simplified dependencies. It enables inference using ONNX models exported from the Python implementation of FoundationPose, making deployment and application highly convenient.

​​Notes:​​ This repository only contains the code for the FoundationPose component. The complete 6D pose estimation pipeline also relies on object masks, which can be generated by algorithms like SAM. For reference implementations and optimized inference of MobileSAM and NanoSAM, please visit EasyDeploy.

Update LOG

[2025.04] Decoupled Register and Track processes; Output poses under mesh coordinates, providing mesh_loader interfaces for external extension. Related PR.

[2025.03] Aligned rendering process with the original Python implementation, supporting rendering without texture input. Related PR.

[2025.03] Added support for Jetson Orin platform with one-click Docker environment setup. See link.

Features

1. Removed complex environment setup and dependency issues from the original project, enabling easy integration with other projects.

2. Implemented encapsulation of the FoundationPose algorithm, ​​supporting dynamic-sized image input​​ for flexible usage. Provided tutorial scripts for generating 3D object models using BundleSDF.

3. 🔥 Supports Jetson Orin development boards (Orin-NX-16GB).

Demo

Test results on public mustard dataset:

foundationpose(fp16) Register test result

foundationpose(fp16) Track test result

Performance on nvidia-4060-8G and i5-12600kf:

nvidia-4060-8G	fps	cpu	gpu
foundationpose(fp16)-Register	2.8	100%	6.5GB
foundationpose(fp16)-Track	220	100%	5.8GB

Performance on jetson-orin-nx-16GB:

jetson-orin-nx-16GB	fps	cpu	mem_total
foundationpose(fp16)-Register	0.6	15%	5.6GB(5.5GB on gpu)
foundationpose(fp16)-Track	100	60%	5.1GB(5.0GB on gpu)

Usage

Environment Setup

1. Clone the repository:

   git clone git@github.com:zz990099/foundationpose_cpp.git
   cd foundationpose_cpp
   git submodule init
   git submodule update

2. Build using Docker:

   cd ${foundationpose_cpp}
   bash easy_deploy_tool/docker/easy_deploy_startup.sh
   # Select `jetson` -> `trt10_u2204`/`trt8_u2204` (`trt8_u2004` not supported)
   bash easy_deploy_tool/docker/into_docker.sh

Model Conversion

1. Download ONNX models from google drive and place them in /workspace/models/.

2. Convert models:

   cd /workspace
   bash tools/cvt_onnx2trt.bash

Build Project

1. Compile the project:

  cd /workspace
  mkdir build && cd build
  cmake -DENABLE_TENSORRT=ON ..
  make -j

Run Demo

Use public Dataset Demo (mustard)

1. Download and extract the dataset to /workspace/test_data/ from here.

2. Run tests:

   cd /workspace/build
   ./bin/simple_tests --gtest_filter=foundationpose_test.test

Custom 3D Model Generation

1. Refer to Generating 3D Models with BundleSDF.

2. Modify paths in /workspace/simple_tests/src/test_foundationpose.cpp for your data and rebuild.

3. Run tests:

   cd /workspace/build
   ./bin/simple_tests --gtest_filter=foundationpose_test.test

4. Results for Register and Track processes will be saved in /workspace/test_data/.

References

• nvidia-isaac-pose-estimation
• FoundationPose
• BundleSDF
• XMem
• EasyDeploy

For any questions, feel free to raise a issue or contact 771647586@qq.com.