A neural framework for the statistical analysis of genus-zero 4D surfaces that deform and evolve overtime. We introduce Dynamic Spherical Neural Surfaces (D-SNS), an efficient and continuous spatiotemporal representation to demonstrate core 4D shape analysis such as spatiotemporal registration, geodesics computation and mean 4D shape estimation framework.
Code:
Dynamic Neural Surfaces for Elastic 4D Shape Representation and Analysis
The code is tested on python=3.11, as well as pytorch=2.5 and torchvision=0.20. Please follow the instructions here to install both PyTorch and TorchVision dependencies. Installing both PyTorch and TorchVision with CUDA support is strongly recommended.
The code also require MATLAB for visualization and computation. Please follow the instructions here to install MATLAB.
We recommend to use conda for installation of all the dependencies. Please follow the command to download the dependencies.
conda env create -f environment.yml
Please download the _pretrained and data folder from the google drive. Please put folder structure in the DSNS-code:
DSNS-code\_pretrained\
DSNS-code\data\
- Dynamic Spherical Neural Surfaces (D-SNS).
python train_dsns.py
- For spatial registration we use Laga et al. codebase please follow the link.
TBA
- Spatiotemporal registration.
python train_time_warp.py
- 4D Geodesics.
TBA
- Co-registration and 4D mean estimation.
TBA
TBA (We plan to release the Jupyter Notebook to run the entire neural framework.)
We showcase the quality of DSNS as heatmap visualization. Similarly, we also highlight spatiotemporal registration visualizations, which are enabled by default. After model training, the visualization will be shown.
This work is made available under the MIT license.