This repository contains the implementation of the paper "Deep Closest Point: Learning Representations for Point Cloud Registration" by Yue Wang et al. The original code is available at Yue Wang's GitHub.

This fork adapts the original Linux-based code to run on a Windows environment. Below are the details of the modifications and the environment in which the code was tested.

• Processor: 13th Gen Intel(R) Core(TM) i7-13700F 2.10 GHz
• RAM: 32.0 GB
• OS: Windows 11 Pro (64-bit operating system, x64-based processor)
• GPU: NVIDIA GeForce RTX 4060 Ti
• CUDA: 11.8

• Python: 3.9.13
• Dependencies:

  pip install torch scipy numpy h5py tqdm tensorboardX

1. OS-Dependent Data Handling:

   • Adjusted code to handle Windows-specific file paths and operations.

2. Deprecated Functions:

   • Replaced deprecated SciPy functions with their updated equivalents.
   • Example:

     r = Rotation.from_dcm(mats[i])

     updated to:

     r = Rotation.from_matrix(mats[i])

3. Optimizer and Scheduler Call Order:

   • Corrected the order of scheduler.step() and optimizer.step() to ensure the learning rate scheduler works as intended.
   • Updated order:

     optimizer.step()
     scheduler.step()

1. Clone the Repository:

   git clone https://github.com/sonwr/dcp.git
   cd dcp

2. Install Dependencies: Make sure you have Python 3.9.13 installed. Then, install the required packages:

   pip install torch scipy numpy h5py tqdm tensorboardX

3. Prepare Data:

   • Ensure your data is correctly formatted and accessible. Modify any paths in the code as necessary to point to your data directories.

4. Run Training:

   • DCP-v1

   python main.py --exp_name=dcp_v1 --model=dcp --emb_nn=dgcnn --pointer=identity --head=svd --eval

   • DCP-v2

   python main.py --exp_name=dcp_v2 --model=dcp --emb_nn=dgcnn --pointer=transformer --head=svd --eval

5. Run Testing:

   • DCP-v1

   python main.py --exp_name=dcp_v1 --model=dcp --emb_nn=dgcnn --pointer=identity --head=svd --eval --model_path=xx/yy

   • DCP-v2

   python main.py --exp_name=dcp_v2 --model=dcp --emb_nn=dgcnn --pointer=transformer --head=svd --eval --model_path=xx/yy

• Ensure your CUDA version matches the version of PyTorch you are using.
• Modify any additional file paths or environment-specific settings as necessary.

The original implementation of this project was done by Yue Wang et al. and is available here. This fork adapts the code to be compatible with Windows OS and includes other minor improvements.

For more details on the paper, refer to the original publication: Deep Closest Point: Learning Representations for Point Cloud Registration.

MIT License