anjanadev96 / NURBS_Diff

NURBS-Diff : A Differentiable NURBS Layer for Machine Learning CAD Applications

NURBS_Diff : A Differentiable NURBS Layer for Machine Learning CAD Applications

NURBS-diff is a differentiable layer that can be run as a standalone layer for CAD applications like curve fitting, surface fitting, surface offseting, and other applications that rely on Non-uniform rational B-splines (NURBS) for representation. NURBS are the current standard for representing CAD geometries, and this work seeks to bridge the gap that currently exists between Deep Learning and Computer-Aided design.
The NURBS-diff layer can also be integrated with other DL frameworks for surface reconstruction to produce accurate rational B-spline surfaces as the output.

Work done at Integrated Design and Engineering Analysis Lab, Iowa State University under Prof. Adarsh Krishnamurthy. Collaborators : Aditya Balu ([e-mail]), Harshil Shah ([e-mail])

Install dependencies

VS 2017 Command Prompt or Anaconda Command Prompt

To install main dependencies for Visual Studio:

1. Open Native x64 VS 2017 command prompt
2. conda create -n 3dlearning
3. conda activate 3dlearning
4. conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch
5. set DISTUTILS_USE_SDK=1 && set PY_VCRUNTIME_REDIST=No thanks && set MSSdk=1
6. Download and unzip CUB (latest release) from https://github.com/NVIDIA/cub/releases
7. set CUB_HOME=path/to/CUB/folder/containing/cmakelists.txt
8. In WSL, edit the following:

• sed -i.bak -e 's/CONSTEXPR_EXCEPT_WIN_CUDA/const/g' /c/tools/miniconda3/envs/test/lib/site-packages/torch/include/torch/csrc/jit/api/module.h
• sed -i.bak -e 's/return \*(this->value)/return \*((type\*)this->value)/g' /c/tools/miniconda3/envs/test/lib/site-packages/torch/include/pybind11/cast.h
• sed -i.bak '/static constexpr Symbol Kind/d' /c/tools/miniconda3/envs/test/lib/site-packages/torch/include/torch/csrc/jit/ir/ir.h

9. pip install "git+https://github.com/facebookresearch/pytorch3d.git"
10. git clone https://github.com/anjanadev96/NURBS_Diff.git
11. cd NURBS_Diff
12. python setup.py develop

• If not already installed via the environment file install NURBS-python by:
  • pip install geomdl

Ubuntu

1. conda create -n 3dlearning
2. conda activate 3dlearning
3. pip install -U fvcore
4. pip install -U iopath
5. conda install -c bottler nvidiacub
6. pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
7. git clone https://github.com/anjanadev96/NURBS_Diff.git
8. python setup.py develop

Examples

Each of the examples can be run using either the CPU version of the code, or the GPU version of the code (available as 'cuda' or 'tc'). \n To run each of the examples, first carry out the build using setup.py.

Curve Fitting

• Code can be found under examples/curve_fitting_on_point_clouds.py
• The layer can be used to fit generic 2D and 3D curves, and point clouds obtained from images.
• To run curve_fitting_on_point_clouds.py, provide a random initialization of input control points, input point cloud and set the number of evaluation points.
• Parameters to vary: degree, number of control points, number of evaluation points.
• Dataset used : Pixel dataset provided under Skelneton challenge.

Surface Fitting

• Code can be found under examples/{surface_fitting.py, nurbs_surface_fitting.py}
• The layer can fit rational and NURBS surfaces.
• Provide input control point grid, number of evaluation points in u, v direction, degree.

Surface Offseting

• Code found under examples for different cases.

Surface reconstruction using Deep Learning

• Splinenet architecture and dataset borrowed from ParSeNet (https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123520256.pdf)
• Trained on 2 NVIDIA Tesla V100s.
• Added support for rational B-splines.

Non-rational B-splines

\n

Rational B-splines

Will be added soon:

• Support for trimmed NURBS surfaces
• Support for automatically learning number of control points
• Dataset for NURBS and trimmed NURBS surfaces