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 found nearest points iteratively.
Additionally, we smooth the calculated mesh with laplacian smoothness term. With laplacian smoothness term, the wireframe is also more neat.
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, see pytorch3d. - For other packages, run
pip install -r requirements.txt
- For the template face model, currently we use a processed version of BFM face model from 3DMMfitting-pytorch, download the BFM09_model_info.mat from 3DMMfitting-pytorch and put it into the ./BFM folder.
- For demo, run
python demo.py
we show demo for NICP mesh2mesh and NICP mesh2pointcloud.
We have two param sets for registration:
milestones = set([50, 80, 100, 110, 120, 130, 140])
stiffness_weights = np.array([50, 20, 5, 2, 0.8, 0.5, 0.35, 0.2])
landmark_weights = np.array([5, 2, 0.5, 0, 0, 0, 0, 0])
This param set is used for registration on fine grained mesh
milestones = set([50, 100])
stiffness_weights = np.array([50, 20, 5])
landmark_weights = np.array([50, 20, 5])
This param set is used for registration on noisy point clouds
Templated Model
You can also use your own templated face model with manually specified landmarks.
Todo
Currently we write some batchwise functions, but batchwise NICP is not supported now. We will support batch NICP in further releases.
