Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces(ICML 2021)

This repository contains the code to reproduce the results from the paper. Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces.

You can find detailed usage instructions for training your own models and using pretrained models below.

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

@inproceedings{NeuralPull,
    title = {Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces},
    author = {Baorui, Ma and Zhizhong, Han and Yu-shen, Liu and Matthias, Zwicker},
    booktitle = {International Conference on Machine Learning (ICML)},
    year = {2021}
}

Surface Reconstruction Demo

Single Image Reconstruction Demo

Installation

First you have to make sure that you have all dependencies in place. The simplest way to do so, is to use anaconda.

You can create an anaconda environment called tensorflow1 using

conda env create -f NeuralPull.yaml
conda activate tensorflow1

Next, for evaluation of the models,compile the extension modules, which are provided by Occupancy Networks. You can do this via

python setup.py build_ext --inplace

To compile the dmc extension, you have to have a cuda enabled device set up. If you experience any errors, you can simply comment out the dmc_* dependencies in setup.py. You should then also comment out the dmc imports in im2mesh/config.py.

Dataset and pretrained model

  1. You can download our preprocessed data and pretrained model.Included in the link:

    –Our pre-train model on ABC and FAMOUS dataset.

    –Preprocessing data of ABC and FAMOUS(sample points and ground truth points).

    –Our reconstruction results.

  2. To make it easier for you to test the code, we have prepared exmaple data in the exmaple_data folder.

Building the dataset

Alternatively, you can also preprocess the dataset yourself. To this end, you have to follow the following steps:

  • Put your own pointcloud files in ‘input_dir’ folder, each pointcloud file in a separate .xyz.npy file.
  • Set an empty folder ‘out_dir’ to place the processed data, note, the folder need to be empty, because this folder will be deleted before the program runs.

You are now ready to build the dataset:

python sample_query_point --out_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --CUDA 0 --dataset other --input_dir ./data/abc_noisefree/04_pts/

Training

You can train a new network from scratch, run

  1. Surface Reconstruction

python NeuralPull.py --data_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --out_dir /data1/mabaorui/AtlasNetOwn/plane_cd_sur/ --class_idx 02691156 --train --dataset shapenet
  1. Single Image Reconstruction

python NeuralPull_SVG.py --data_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --out_dir /data1/mabaorui/AtlasNetOwn/plane_cd_sur/ --class_idx 02691156 --train --class_name plane
  1. Train the dataset yourself

python NeuralPull.py --data_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --out_dir /data1/mabaorui/AtlasNetOwn/plane_cd_sur/ --class_idx 02691156 --train --dataset other

Evaluation

For evaluation of the models and generation meshes using a trained model, use

  1. Surface Reconstruction

python NeuralPull.py --data_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --out_dir /data1/mabaorui/AtlasNetOwn/plane_cd_sur/ --class_idx 02691156 --dataset shapenet
  1. Single Image Reconstruction

python NeuralPull_SVG.py --data_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --out_dir /data1/mabaorui/AtlasNetOwn/plane_cd_sur/ --class_idx 02691156 --class_name plane
  1. Evaluation the dataset yourself

python NeuralPull.py --data_dir /data1/mabaorui/AtlasNetOwn/data/plane_precompute_2/ --out_dir /data1/mabaorui/AtlasNetOwn/plane_cd_sur/ --class_idx 02691156 --dataset other

Script Parameters Explanation

Parameters Description
train train or test a network.
data_dir preprocessed data.
out_dir store network parameters when training or to load pretrained network parameters when testing.
class_idx the class to train or test when using shapenet dataset, other dataset, default.
class_name the class to train or test when using shapenet dataset, other dataset, default.
dataset shapenet,famous,ABC or other(your dataset)

Pytorch Implementation of Neural-Pull

Notably, the code in Pytorch implementation is not released by the official lab, it is achieved by @wzxshgz123’s diligent work. His intention is only to provide references to researchers who are interested in Pytorch implementation of Neural-Pull. There is no doubt that his unconditional dedication should be appreciated.

GitHub

https://github.com/mabaorui/NeuralPull