DeepMetaHandles (CVPR2021 Oral)

DeepMetaHandles is a shape deformation technique. It learns a set of meta-handles for each given shape. The disentangled meta-handles factorize all the plausible deformations of the shape, while each of them corresponds to an intuitive deformation direction. A new deformation can then be generated by the "linear combination" of the meta-handles. Although the approach is learned in an unsupervised manner, the learned meta-handles possess strong interpretability and consistency.

DeepMetaHandles-1

Environment setup

  1. Create a conda environment by conda env create -f environment.yml.
  2. Build and install torch-batch-svd.

Demo

  1. Download data/demo and checkpoints/chair_15.pth from here and place them in the corresponding folder. Pre-processed demo data contains the manifold mesh, sampled control point, sampled surface point cloud, and corresponding biharmonic coordinates.
  2. Run src/demo_target_driven_deform.py to deform a source shape to match a target shape.
  3. Run src/demo_meta_handle.py to generate deformations along the direction of each learned meta-handle.

Train

  1. Download data/chair from here and place them in the corresponding folder.
  2. Run the visdom server. (We use visdom to visualize the training process.)
  3. Run src/train.py to start training.

Note: For different categories, you may need to adjust the number of meta-handles. Also, you need to tune the weights for the loss functions. Different sets of weights may produce significantly different results.

Pre-process your own data

  1. Compile codes in data_preprocessing/.
  2. Build and run manifold to convert your meshes into watertight manifolds.
  3. Run data_preprocessing/normalize_bin to normalize the manifold into a unit bounding sphere.
  4. Build and run fTetWild to convert your manifolds into tetrahedral meshes. Please use --output xxx.mesh option to generate the .mesh format tet mesh. Also, you will get a xxx.mesh__sf.obj for the surface mesh. We will use xxx.mesh and xxx.mesh__sf.obj to calculate the biharmonic weights. We will only deform xxx.mesh__sf.obj later.
  5. Run data_preprocessing/sample_key_points_bin to sample control points from xxx.mesh__sf.obj. We use the FPS algorithm over edge distances to sample the control points.
  6. Run data_preprocessing/calc_weight_bin to calculate the bihrnomic weights. It takes xxx.mesh, xxx.mesh__sf.obj, and the control point file as input, and will output a text file containing the weight matrix for the vertices in xxx.mesh__sf.obj.
  7. Run data_preprocessing/sample_surface_points_bin to sample points on the xxx.mesh__sf.obj and calculate the corresponding biharmonic weights for the sampled point cloud.
  8. In our training, we remove those shapes (about 10%) whose biharmonic weight matrix contains elements that are smaller than -1.5 or greater than 1.5. We find that this can help us to converge faster.
  9. To reduce IO time during training, you may compress the data into a compact form and load them to the memory.

Citation

If you find our work useful, please consider citing our paper:

@article{liu2021deepmetahandles,
  title={DeepMetaHandles: Learning Deformation Meta-Handles of 3D Meshes with Biharmonic Coordinates},
  author={Liu, Minghua and Sung, Minhyuk and Mech, Radomir and Su, Hao},
  journal={arXiv preprint arXiv:2102.09105},
  year={2021}
}

GitHub

https://github.com/Colin97/DeepMetaHandles