GPU Accelerated Non-rigid ICP for surface registration

Introduction

Preivous Non-rigid ICP algorithm is usually implemented on CPU, and needs to solve sparse least square problem, which is time consuming. In this repo, we implement a pytorch version NICP algorithm based on paper Amberg et al. Detailedly, we leverage the AMSGrad to optimize the linear regresssion, and then find nearest points iteratively. Additionally, we smooth the calculated mesh with laplacian smoothness term. With laplacian smoothness term, the wireframe is also more neat. Also, we support batchwise registration, in each batch, we can registering different meshes with the same template.

---

Quick Start

Install

We use python3.8 and cuda10.2 for implementation. The code is tested on Ubuntu 20.04.

• The pytorch3d cannot be installed directly from pip install pytorch3d, for the installation of pytorch3d, please follow the instructions from pytorch3d.
• For other packages, run

pip install -r requirements.txt

Demo

• For the template face model, currently we use a processed version of BFM face model from 3DMMfitting-pytorch. Please download the BFM09_model_info.mat from 3DMMfitting-pytorch and put it into the ./BFM folder.
• For demo, run

python demo_nicp.py

Then the NICP from mesh2mesh and NICP from mesh2pointcloud will be conducted. We have two configuration files ./config/fine_grain.json and ./config/coarse_grain.json, which is used respectively for mesh2mesh and mesh2pointcloud registration. The ./config/fine_grain.json has less stiffness constraints, which is suitable for registration on accurate scan. The ./config/coarse_grain.json has less stiffness constraints, which is suitable for registration on noisy pointclouds.

• We also provide a demo on 4D depth data, we provide a demo in demo_4D.py, also, we upload the example depth data in test_data/depth_wuhz.mp4, this data is collected from realsense L515. On average, each frame takes 2 seconds for processing.

Mesh2Mesh

Mesh2Pointcloud

Notes

The target mesh/pointcloud should face towards z-axis as shown in the following picture. Meanwhile, the mesh/pointcloud must be normalized with utils.normalize_mesh or utils.normalize_pcl before fed into the nicp function.

Direction

Batchwise NICP

Currently we write some batchwise functions, but batchwise NICP is not supported now. We will support batch NICP in further releases.