A Survey and Benchmark of Automatic Surface Reconstruction from Point Clouds

Data and evaluation code for the paper A Survey and Benchmark of Automatic Surface Reconstruction from Point Clouds (arXiv).

alt text

:floppy_disk: Datasets

Berget et al.

The watertight meshes and scanned point clouds used in our paper can be downloaded here.

ModelNet10

The watertight ModelNet10 models can be downloaded here on Zenodo.
The scanned point clouds used in our paper can be downloaded here on Zenodo. The dataset also includes evaluation data and training data for ConvONet, Points2Surf, Shape As Points, POCO and DGNN.

ShapeNetv1 (13 class subset of Choy et al.)

The watertight ShapeNet models can be downloaded here (provided by the authors of ONet).
To produce the scanned point clouds used in our paper follow the instructions below.
Training and evaluation data for ShapeNet can be downloaded here (provided by the authors of ONet).

:chart_with_upwards_trend: Evaluate your surface reconstruction method

1) Clone and install this repository with

git clone https://github.com/raphaelsulzer/dsr-benchmark.git
cd dsr-benchmark
bash install.sh   # create a conda environment called dsr including all python dependencies
conda activate dsr

2) Clone and install mesh-tools to your CPP_DIR.

3) Download and unzip the three datasets above (Berger, ModelNet, ShapeNet) to your DATASET_DIR.

4) Test if everything is installed correctly by running.

python test/test_benchmark.py --cpp_dir CPP_DIR --dataset_dir DATASET_DIR

5) Apply your reconstruction algorithm and evaluate your reconstructions.

from dsrb.datasets import Berger, ModelNet10, ShapeNet
from dsrb.eval import MeshEvaluator

me = MeshEvaluator()

####################################################
### Berger et al. dataset (Experiments 4, 6 & 7) ###
####################################################
dataset = Berger()
# get the MVS scans used in Experiment 4 (learning-based)
models = dataset.get_models(scan_configuration="mvs")
# or, get the LiDAR scans used in Experiment 6 (optimization-based)
models = dataset.get_models(scan_configuration=["0","1","2","3","4"])
for model in models:
    input_data = model["scan"]
    YOUR_RECONSTRUCTION_ALGORITHM(
        points=input_data["points"],
        normals=input_data["normals"], # optional
        sensor_position=input_data["sensor_position"], # optional
        # # or, if your algorithm inputs a .ply file with points and normals
        # infile=model["scan_ply"],
        outfile=model["output"]["surface"].format("NAME_OF_YOUR_ALGORITHM"))
# evaluate your reconstructions 
me.eval(models,outpath=dataset.path,method="NAME_OF_YOUR_ALGORITHM")

###############################################
### ModelNet10 dataset (Experiments 3 & 5 ) ###
###############################################
dataset = ModelNet10()
models = dataset.get_models()
models_train = models["train"]
for model in models["test"]:
    input_data = model["scan"]
    YOUR_RECONSTRUCTION_ALGORITHM(
        points=input_data["points"],
        normals=input_data["normals"], # optional
        sensor_position=input_data["sensor_position"], # optional
        outfile=model["output"]["surface"].format("NAME_OF_YOUR_ALGORITHM"))
# evaluate your reconstructions
me.eval(models,outpath=dataset.path,method="NAME_OF_YOUR_ALGORITHM")

############################################
### ShapeNet dataset (Experiments 1 - 5) ###
############################################
dataset = ShapeNet()
dataset.setup()
# the scanned point clouds for ShapeNet are not included in the downloaded dataset
# generate the 10k scans with outliers and noise (Experiment 2) 
models = dataset.get_models(splits="test")
dataset.scan(scan_configuration="6") # takes around 4h
# generate the 3k scans with noise
dataset.clear()
models = dataset.get_models()
dataset.scan(scan_configuration="4") # takes around 4h
# now get either scan 4 (3k noise) or scan 6 (10k noise and outliers)
dataset.clear()
models = dataset.get_models(scan_configuration="4")
models_train = models["train"]
models_validation = models["val"]
for model in models["test"]:
    input_data = model["scan"]
    YOUR_RECONSTRUCTION_ALGORITHM(
        points=input_data["points"],
        normals=input_data["normals"], # optional
        sensor_position=input_data["sensor_position"], # optional
        outfile=model["output"]["surface"].format("NAME_OF_YOUR_ALGORITHM"))
# evaluate your reconstructions
me.eval(models,outpath=dataset.path,method="NAME_OF_YOUR_ALGORITHM")

:book: Citation

If you find the code or data in this repository useful, please consider citing

@misc{sulzer2023dsr,
  doi = {10.48550/ARXIV.2301.13656},
  url = {https://arxiv.org/abs/2301.13656},
  author = {Sulzer, Raphael and Landrieu, Loic and Marlet, Renaud and Vallet, Bruno},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), Computational Geometry (cs.CG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {A Survey and Benchmark of Automatic Surface Reconstruction from Point Clouds},
  publisher = {arXiv},
  year = {2023},
}

raphaelsulzer / dsr-benchmark

readme