ESM_ICP

ESM_ICP is a C++ implementation of the Exponential Similarity Matrix Iterative Closest Point (ESM-ICP) algorithm for point cloud registration. This repository provides tools for aligning 3D point clouds using the ESM-ICP method.

Features

• Implementation of the ESM-ICP algorithm for precise point cloud registration.
• Visualization tools for analyzing registration results.
• Modular code structure for easy integration and extension.

📄 Reference

If you use this code or dataset, please cite the following paper:

A. Singandhupe,Sanket Lokhande, H. La, "Registration of 3D Point Sets Using Exponential-based Similarity Matrix", 2025.

📄 Read the paper (PDF)

Getting Started

Prerequisites

• C++17 compatible compiler
• Eigen library
• PCL (Point Cloud Library) (optional, for visualization)

Building the Project

1. Clone the repository:

   git clone --recurse-submodules https://github.com/aralab-unr/ESM_ICP.git
   cd ESM_ICP
   mkdir build
   cd build
   cmake ..
   make
   

2. NOTE: Create folder res/ and saved_pcd/ in you cloned repository (this is for logging purposes), other wise the program crashes OR in the code ESM_ICP_demo_Viewer.cpp comment out the lines which are logging in data in these folder.

3. Usage

     ./ESM_ICP_deo_Viewer <source.pcd> <method> <leaf_size> <outlier_percent>
     for example,
    ./ESM_ICP_demo_Viewer ../../datasets/ModelNet40_pcd/stool/test/stool_0092.pcd icp 0.009 0.0f
   
   <method> can be icp, gicp, icp_nl, point2plane, ndt, goicp, deepgmr, dcp, pointnetlk and rpmnet.
   <leaf_size> set the leaf size to downsample the point clouds.
   <outlier_percent> denotes how much percent of points do you want the source to be affected by noise.
   

4. Hit 'SPACE' in the viewer. Every time you hit space it will trigger 40 iterations of ESM-ICP in the left side. For the right side it will trigger 40 iterations only for ICP/ICP-NL/GICP/point2plane. For the methods like Go-ICP/other deep learning methods, it will perform only 1 iteration.

5. If you want to use GO-ICP go to the Go-ICP/ folder and do mkdir build, cd build, cmake .. and make.

6. IMP NOTES:

   • The 'method' is used only for comparison with ESM-ICP. In the viewer's left view, the ouput is always from ESM-ICP and in the right view the output is from the method you pick.
   • The code randomly generates a transformation using the function generateRandomTransformation() and transfroms the loaded point point cloud.
   • ESM-ICP performs alignment on the original and the transformed point cloud along with other methods for comparison.
   • Honestly, the actual implementation is in the function "Eigen::Matrix4d computeOptimalTransformation(const PointCloudT::Ptr& source, const PointCloudT::Ptr& target, const std::vector& correspondences)", the remaining are just codes for comparison with other methods and can be safely ignored.
   • just focus on the ESM_ICP_demo_Viewer.cpp , the remaining cpp files are for testing purposes only.
   • GPU implementation is still under process.