MetaSDF: Meta-learning Signed Distance Functions

Project Page | Paper | Data

Vincent Sitzmann*, Eric Ryan Chan*, Richard Tucker, Noah Snavely
Gordon Wetzstein
*denotes equal contribution

This is the official implementation of the paper "MetaSDF: Meta-Learning Signed Distance Functions".

In this paper, we show how we may effectively learn a prior over implicit neural representations using gradient-based meta-learning.

While in the paper, we show this for the special case of SDFs with the ReLU nonlinearity, this works formidably well with other types of neural implicit representations - such as our work "SIREN"!

We show you how in our Colab notebook:

DeepSDF

A large part of this codebase (directory "3D") is based on the code from the terrific paper "DeepSDF" - check them out!

Get started

If you only want to experiment with MetaSDF, we have written a colab that doesn't require installing anything, and goes through a few other interesting properties of MetaSDF as well - for instance, it turns out you can train SIREN to fit any image in only just three gradient descent steps!

If you want to reproduce all the experiments from the paper, you can then set up a conda environment with all dependencies like so:

conda env create -f environment.yml
conda activate metasdf

3D Experiments

Dataset Preprocessing

Before training a model, you'll first need to preprocess the training meshes. Please follow the preprocessing steps used by DeepSDF if using ShapeNet.

Define an Experiment

Next, you'll need to define the model and hyperparameters for your experiment. Examples are given in 3D/curriculums.py, but feel free to make modifications. Although not present in the original paper, we've included some curriculums with positional encodings and smaller models. These generally perform on par with the original models but require much less memory.

Train a Model

After you've preprocessed your data and have defined your curriculum, you're ready to start training! Navigate to the 3D/scripts directory and run

python run_train.py <curriculum name>.

If training is interupted, pass the flag --load flag to continue training from where you left off.

You should begin seeing printouts of loss, with a summary at every epoch. Checkpoints and Tensorboard summaries are saved to the 'output_dir' directory, as defined in your curriculum. We log raw loss, which is either the composite loss or L1 loss, depending on your experiment definition, as well as a 'Misclassified Percentage'. The 'Misclassified Percentage' is the percentage of samples that the model incorrectly classified as inside or outside the mesh.

Reconstructing Meshes

After training a model, recontruct some meshes using

python run_reconstruct.py <curriculum name> --checkpoint <checkpoint file name>.

The script will use the 'test_split' as defined in the curriculum.

Evaluating Reconstructions

After reconstructing meshes, calculate Chamfer Distances between reconstructions and ground-truth meshes by running

python run_eval.py <reconstruction dir>.

Torchmeta

We're using the excellent torchmeta to implement hypernetworks.

Citation

If you find our work useful in your research, please cite:

       @inproceedings{sitzmann2019metasdf,
            author = {Sitzmann, Vincent
                      and Chan, Eric R.
                      and Tucker, Richard
                      and Snavely, Noah
                      and Wetzstein, Gordon},
            title = {MetaSDF: Meta-Learning Signed
                     Distance Functions},
            booktitle = {Proc. NeurIPS},
            year={2020}
       }

Contact

If you have any questions, please feel free to email the authors.