Awesome

GenSDF: Two-Stage Learning of Generalizable Signed Distance Functions

Paper | Supplement | Project Page | Teaser

promo_gif

This repository contains the official implementation of GenSDF: Two-Stage Learning of Generalizable Signed Distance Functions Gene Chou, Ilya Chugunov, Felix Heide NeurIPS 2022

If you find our code or paper useful, please consider citing

@inproceedings{chou2022gensdf,
title={GenSDF: Two-Stage Learning of Generalizable Signed Distance Functions},
author={Chou, Gene and Chugunov, Ilya and Heide, Felix},
booktitle={Proc. of Neural Information Processing Systems (NeurIPS)},
year={2022}
}

Additionally, this repository is designed for reproduction and incorporation of SDF and Occupancy methods. Currently includes GenSDF, DeepSDF, Convolutional Occupancy Networks, and NeuralPull. Note that these methods except for GenSDF are not necessarily optimized and may require minor tweaks (e.g., hyperparameter changes) for optimal performance; we encourage users to validate with original implementations. We provide links and references below.

SDF lib
  ├── config  
  │   └── // folders for checkpoints and training configs
  ├── data  
  │   └── // folders for data; in csv format
  ├── model  
  │   ├── gensdf
  │   │   └── model.py
  │   ├── deepsdf
  │   │    └── model.py 
  │   ├── archs
  │   │    └── encoders
  │   │        └── // pointnet, auto-decoder...etc
  │   │    └── decoders
  │   │        └── // 8-layer MLP, SIREN...etc
  │   └── ... // add your own models
  ├── dataloader  
  │   └── // dataloaders for labeled and unlabeled data are provided
  ├── utils  
  │   └── // scripts for reconstruction and evaluation
  ├── environment.yml  // package requirements
  ├── train.py  // script for training; change dataloader here
  ├── test.py  // script for testing given a defined testing dataloader
  └── test_single.py  // script for testing given a single input point cloud

Installation

We recommend creating an anaconda environment using our provided environment.yml:

conda env create -f environment.yml
conda activate sdf

Note: you might need to install torch-scatter manually following the official instruction:

pip install torch-scatter

Demo (for GenSDF)

Run test_single.py. Load a point cloud in csv format (see data/acronym) or ply format (see data/ycb) and visualize reconstructions.

# give file path and output directory
python test_single.py -f data/acronym/Bear/52f9a230c400bddd55330ea4156a425f/sdf_data.csv -o .

Dataset

For training, we first recenter and normalize input data. We also apply these steps for an input point cloud in test_single.py. See supplement for an explanation. For training, we preprocess all files and store results in csv format. Each csv file corresponds to one object, and each line represents a coordinate followed by its signed distance value. See data/acronym for examples.

For labeled data (i.e., meshes), we recommend first recentering and normalizing the object, then sampling point clouds and query points. The ground truth signed distance values for all points can be calculated from the mesh and stored into csv files.

We train using the Acronym dataset. Place your dataset in the data folder following the current structure, and provide the train/test split in the config spec files.

Training

We provide pretrained models for GenSDF. Users are also welcome to train from scratch using the following commands

python train.py -e config/gensdf/semi -b 64 -r last  # remove -r to train from scratch
python train.py -e config/deepsdf -b 64 -w 12        # -b for batch size, -w for workers

Feel free to modify specs.json files in config. One can also mix and match encoders and decoders that are provided.

Testing

For testing single objects, see the Demo section above. For testing on multiple objects, define a dataloader and run the test.py file

python test.py -e config/gensdf/semi -r last

References

We adapt code from DeepSDF: https://github.com/facebookresearch/DeepSDF Convolutional Occupancy Networks: https://github.com/autonomousvision/convolutional_occupancy_networks NeuralPull: https://github.com/bearprin/neuralpull-pytorch/tree/history-2