Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time

DeepIM: Deep Iterative Matching for 6D Pose Estimation

Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang and Dieter Fox. In ECCV, 2018. arXiv, project page

This is an official MXNet implementation mainly developed and maintained by Yi Li and Gu Wang.

News (2020-12-04): A PyTorch implementation of DeepIM by Yu Xiang has been released (here)!

Citing DeepIM

If you find DeepIM useful in your research, please consider citing:

title     = {DeepIM: Deep Iterative Matching for 6D Pose Estimation},
author    = {Yi Li and Gu Wang and Xiangyang Ji and Yu Xiang and Dieter Fox},
booktitle = {European Conference on Computer Vision (ECCV)},
year      = {2018}

Overall Framework

Network Structure

Zoom In Operation

Main Results


Occlusion LINEMOD

Unseen Objects from ModelNet

The red and green lines represent the edges of 3D model projected from the initial poses and our refined poses respectively.

Requirements: Software

  1. Python 2.7. We recommend using Anaconda. (python 3.x should also be OK.)

  2. GLFW for OpenGL: sudo apt-get install libglfw3-dev libglfw3 (on Ubuntu 16.04)

  3. Python packages might missing:

    conda install scipy
    pip install Cython
    pip install opencv-python
    pip install easydict
    pip install pyyaml
    pip install tqdm


    pip install pyopengl packaging appdirs pyopengl triangle cython glfw
    # clone the lastest glumpy (there is a bug in the pip version)
    git clone
    cd glumpy
    pip install .
  4. MXNet from the official repository.

    Option 1: Use the prebuilt version following the installation guide..

    nvcc --version
    pip install mxnet-cu90 # (change to your cuda version)

    Option 2. Build MXNet from the source following the official manual:

    2.1 Clone MXNet and checkout to MXNet@(commit fc9e70b) by

    git clone --recursive
    cd mxnet
    git checkout fc9e70b (optional)
    git submodule update (optional)

    or use the latest master directly (code is tested under mxnet 1.2.0).

    2.2 Compile MXNet

    cd ${MXNET_ROOT}
    make -j $(nproc) USE_OPENCV=1 USE_BLAS=openblas USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1

    2.3 Install the MXNet Python binding by

    Note: If you will actively switch between different versions of MXNet, please follow 2.4

    cd python
    sudo python install

    2.4 For advanced users, you may put your Python packge into ./external/mxnet/$(YOUR_MXNET_PACKAGE), and modify MXNET_VERSION in ./experiments/deepim/cfgs/*.yaml to $(YOUR_MXNET_PACKAGE). Thus you can switch among different versions of MXNet quickly.

  5. Use tensorboard to visualize loss:

    Install mxboard following

    pip install mxboard

Requirements: Hardware

Any NVIDIA GPUs with at least 4GB memory should be OK.


  1. Clone the DeepIM repository, and we'll call the directory that you cloned mx-DeepIM as ${DeepIM_ROOT}.
git clone
  1. Initialize DeepIM:

    2.1 In the root directory of DeepIM, run sh to initialize the DeepIM project. (Note: For python3, need to install pytorch first to jit compile flow_c module.)

    2.2 (Optional) Delete the data folder and link (i.e. ln -sf) the root folder of data to ./data.

Preparation for Training & Testing

  1. Prepare datasets, see ./toolkit/ and for details.

    The datasets should be put in folder:

  2. Please download FlowNet model manually from Google Drive or Baidu NetDisk (password: shga), and put it under folder ./model. Make sure it looks like this:



  1. All of our experiment settings (GPU, dataset, etc.) are kept in yaml config files at folder ./experiments/deepim/cfgs

  2. To perform experiments, run the python scripts with the corresponding config file as input. For example, to train and test DeepIM models with pre-trained FlowNet, use the following command

    python experiments/deepim/ --cfg experiments/deepim/cfgs/your_cfg.yaml --gpus 0,1,2,3

    A cache folder would be created automatically to save the model and the log under output/deepim/. or you can just run the script like


    to train and test the model only on category ape or run


    to train and test the model on all categories.

    Trained weights for LINEMOD and Occlusion LINEMOD can be found here, Google Drive.

  3. Please find more details in config files and in our code.


Code has been tested under:

  • Ubuntu 14.04/16.04 with 4 GTX 1080Ti GPUs or a single GTX 1070 GPU


Deep Iterative Matching for 6D Pose Estimation







No packages published