[Paper] [Website]

This is the official implementation of our paper to be appeared in CVPR 2024 (Highlight)

Contributors: Bowen Wen, Wei Yang, Jan Kautz, Stan Birchfield

We present FoundationPose, a unified foundation model for 6D object pose estimation and tracking, supporting both model-based and model-free setups. Our approach can be instantly applied at test-time to a novel object without fine-tuning, as long as its CAD model is given, or a small number of reference images are captured. We bridge the gap between these two setups with a neural implicit representation that allows for effective novel view synthesis, keeping the downstream pose estimation modules invariant under the same unified framework. Strong generalizability is achieved via large-scale synthetic training, aided by a large language model (LLM), a novel transformer-based architecture, and contrastive learning formulation. Extensive evaluation on multiple public datasets involving challenging scenarios and objects indicate our unified approach outperforms existing methods specialized for each task by a large margin. In addition, it even achieves comparable results to instance-level methods despite the reduced assumptions.

🤖 For ROS version, please check Isaac ROS Pose Estimation, which enjoys TRT fast inference and C++ speed up.

🥇 No. 1 on the world-wide BOP leaderboard (as of 2024/03) for model-based novel object pose estimation.

Robotic Applications:

AR Applications:

Results on YCB-Video dataset:

@InProceedings{foundationposewen2024,
author        = {Bowen Wen, Wei Yang, Jan Kautz, Stan Birchfield},
title         = {{FoundationPose}: Unified 6D Pose Estimation and Tracking of Novel Objects},
booktitle     = {CVPR},
year          = {2024},
}

If you find the model-free setup useful, please also consider cite:

@InProceedings{bundlesdfwen2023,
author        = {Bowen Wen and Jonathan Tremblay and Valts Blukis and Stephen Tyree and Thomas M\"{u}ller and Alex Evans and Dieter Fox and Jan Kautz and Stan Birchfield},
title         = {{BundleSDF}: {N}eural 6-{DoF} Tracking and {3D} Reconstruction of Unknown Objects},
booktitle     = {CVPR},
year          = {2023},
}

1. Download all network weights from here and put them under the folder weights/. For the refiner, you will need 2023-10-28-18-33-37. For scorer, you will need 2024-01-11-20-02-45.

2. Download demo data and extract them under the folder demo_data/

3. [Optional] Download our large-scale training data: "FoundationPose Dataset"

4. [Optional] Download our preprocessed reference views here in order to run model-free few-shot version.

cd docker/
docker pull wenbowen123/foundationpose && docker tag wenbowen123/foundationpose foundationpose  # Or to build from scratch: docker build --network host -t foundationpose .
bash docker/run_container.sh

If it's the first time you launch the container, you need to build extensions.

bash build_all.sh

Later you can execute into the container without re-build.

docker exec -it foundationpose bash

• Setup conda environment

# create conda environment
conda create -n foundationpose python=3.9

# activate conda environment
conda activate foundationpose

# Install Eigen3 3.4.0 under conda environment
conda install conda-forge::eigen=3.4.0
export CMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH:/eigen/path/under/conda"

# install dependencies
python -m pip install -r requirements.txt

# Install NVDiffRast
python -m pip install --quiet --no-cache-dir git+https://github.com/NVlabs/nvdiffrast.git

# Kaolin (Optional, needed if running model-free setup)
python -m pip install --quiet --no-cache-dir kaolin==0.15.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.0.0_cu118.html

# PyTorch3D
python -m pip install --quiet --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py39_cu118_pyt200/download.html

# Build extensions
CMAKE_PREFIX_PATH=$CONDA_PREFIX/lib/python3.9/site-packages/pybind11/share/cmake/pybind11 bash build_all_conda.sh

The paths have been set in argparse by default. If you need to change the scene, you can pass the args accordingly. By running on the demo data, you should be able to see the robot manipulating the mustard bottle. Pose estimation is conducted on the first frame, then it automatically switches to tracking mode for the rest of the video. The resulting visualizations will be saved to the debug_dir specified in the argparse. (Note the first time running could be slower due to online compilation)

python run_demo.py

Feel free to try on other objects (no need to retrain) such as driller, by changing the paths in argparse.

For this you first need to download LINEMOD dataset and YCB-Video dataset.

To run model-based version on these two datasets respectively, set the paths based on where you download. The results will be saved to debug folder

python run_linemod.py --linemod_dir /mnt/9a72c439-d0a7-45e8-8d20-d7a235d02763/DATASET/LINEMOD --use_reconstructed_mesh 0

python run_ycb_video.py --ycbv_dir /mnt/9a72c439-d0a7-45e8-8d20-d7a235d02763/DATASET/YCB_Video --use_reconstructed_mesh 0

To run model-free few-shot version. You first need to train Neural Object Field. ref_view_dir is based on where you download in the above "Data prepare" section. Set the dataset flag to your interested dataset.

python bundlesdf/run_nerf.py --ref_view_dir /mnt/9a72c439-d0a7-45e8-8d20-d7a235d02763/DATASET/YCB_Video/bowen_addon/ref_views_16 --dataset ycbv

Then run the similar command as the model-based version with some small modifications. Here we are using YCB-Video as example:

python run_ycb_video.py --ycbv_dir /mnt/9a72c439-d0a7-45e8-8d20-d7a235d02763/DATASET/YCB_Video --use_reconstructed_mesh 1 --ref_view_dir /mnt/9a72c439-d0a7-45e8-8d20-d7a235d02763/DATASET/YCB_Video/bowen_addon/ref_views_16

• For more recent GPU such as 4090, refer to this.

• For setting up on Windows, refer to this.

• If you are getting unreasonable results, check this

Our training data include scenes using 3D assets from GSO and Objaverse, rendered with high quality photo-realism and large domain randomization. Each data point includes RGB, depth, object pose, camera pose, instance segmentation, 2D bounding box. [Google Drive].

• To parse the camera params including extrinsics and intrinsics

  with open(f'{base_dir}/camera_params/camera_params_000000.json','r') as ff:
    camera_params = json.load(ff)
  world_in_glcam = np.array(camera_params['cameraViewTransform']).reshape(4,4).T
  cam_in_world = np.linalg.inv(world_in_glcam)@glcam_in_cvcam
  world_in_cam = np.linalg.inv(cam_in_world)
  focal_length = camera_params["cameraFocalLength"]
  horiz_aperture = camera_params["cameraAperture"][0]
  vert_aperture = H / W * horiz_aperture
  focal_y = H * focal_length / vert_aperture
  focal_x = W * focal_length / horiz_aperture
  center_y = H * 0.5
  center_x = W * 0.5
  
  fx, fy, cx, cy = focal_x, focal_y, center_x, center_y
  K = np.eye(3)
  K[0,0] = fx
  K[1,1] = fy
  K[0,2] = cx
  K[1,2] = cy
  

Due to the legal restrictions of Stable-Diffusion that is trained on LAION dataset, we are not able to release the diffusion-based texture augmented data, nor the pretrained weights using it. We thus release the version without training on diffusion-augmented data. Slight performance degradation is expected.

We would like to thank Jeff Smith for helping with the code release; NVIDIA Isaac Sim and Omniverse team for the support on synthetic data generation; Tianshi Cao for the valuable discussions. Finally, we are also grateful for the positive feebacks and constructive suggestions brought up by reviewers and AC at CVPR.

The code and data are released under the NVIDIA Source Code License. Copyright © 2024, NVIDIA Corporation. All rights reserved.

For questions, please contact Bowen Wen.