Skip to content


Repository files navigation

POCO: Point Convolution for Surface Reconstruction

by: Alexandre Boulch and Renaud Marlet

Computer Vision and Pattern Recognition, CVPR, 2022

Paper    PDF    Supp    Arxiv



Implicit neural networks have been successfully used for surface reconstruction from point clouds. However, many of them face scalability issues as they encode the isosurface function of a whole object or scene into a single latent vector. To overcome this limitation, a few approaches infer latent vectors on a coarse regular 3D grid or on 3D patches, and interpolate them to answer occupancy queries. In doing so, they loose the direct connection with the input points sampled on the surface of objects, and they attach information uniformly in space rather than where it matters the most, i.e., near the surface. Besides, relying on fixed patch sizes may require discretization tuning. To address these issues, we propose to use point cloud convolutions and compute latent vectors at each input point. We then perform a learning-based interpolation on nearest neighbors using inferred weights. Experiments on both object and scene datasets show that our approach significantly outperforms other methods on most classical metrics, producing finer details and better reconstructing thinner volumes.


    author    = {Boulch, Alexandre and Marlet, Renaud},
    title     = {POCO: Point Convolution for Surface Reconstruction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {6302-6314}


The code was used with Ubuntu, Python 3.7.10, Cuda 11.1 and Pytorch 1.8.1

Installing APT packages

apt-get install libgl1-mesa-glx libopenblas-dev

Creating a conda environement

If you want to run the code in a conda environment, you can create one with Python 3.7.10 and CudaToolKit 11.1 using:

conda env create -f environment.yml
conda activate poco

Installing dependencies

pip install -r requirements.txt

Building the triangle hash module (from ConvONet)

This module is used for score computation

CC=gcc CXX=gcc python build_ext --inplace
python build_ext --inplace


ShapeNet (Occupancy Network pre-processing)

We use the ShapeNet dataset as pre-processed by Occupancy Networks. Please refer to original repository for downloading the data.

It should be placed in the folder data/ShapeNet.

SyntheticRooms (Occupancy Network pre-processing)

We use the SyntheticRooms dataset as created by Occupancy Networks. Please refer to original repository for downloading the data.

It should be placed in the folder data/synthetic_room_dataset.

ABC / RealWorld / Famous and Thingi10k

They should be placed in the folders: data/3d_shapes_abc, data/3d_shapes_abc_training, data/3d_shapes_famous, data/3d_shapes_real_world and data/3d_shapes_thingi10k.


Download the SceneNet dataset.

Watertight models

We use the code from for watertight model creation. Please download and compile it.

To create the watertight models, create a destination directory data/SceneNet, copy-paste the python script datasets/ into the directory and run it with python You will need to modify the raw_data_dir and the manifold_code_dir to match your data and code locations.

Creation of the point clouds

You can use the script datasets/ You will also need to modify the paths.

Precomputed meshes and point clouds

For easy use of the code, we provide precomputed meshes, and point clouds SceneNet20, SceneNet100, SceneNet500 and SceneNet1000.


The watertight models should be in data/SceneNet.

The points are located in data/SceneNet20, data/SceneNet100, data/SceneNet1000 and data/SceneNet500.


# train on ShapeNet with 3k points, noise and without normals 
python --config configs/config_shapenet.yaml 

# train on ShapeNet with 10k points, no noise and normals
python --config configs/config_shapenet.yaml --normals True --random_noise 0 --experiment_name Normals

# train on SyntheticRooms
python --config configs/config_synthetic.yaml

# train on ABC with 3k points
python --config configs/config_abc.yaml --experiment_name 3k

# train on ABC with 10k points
python --config configs/config_abc.yaml --experiment_name 10k --manifold_points 10000 --training_batch_size 8



python --config results/ShapeNet_None_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --gen_resolution_global 128


python --config results/SyntheticRooms_None_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --gen_resolution_global 256 --num_mesh 1

ABC, RealWorld, Famous and Thingi10k

# Model trained with 3k points
python --config results/ABC_3k_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --dataset_name DATASET_NAME --dataset_root data/DATASET_DIR --gen_resolution_global 256

python --config results/ABC_3k_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --dataset_name DATASET_NAME --dataset_root data/DATASET_DIR --gen_resolution_global 256 --manifold_points -1 --gen_subsample_manifold 3000 --gen_subsample_manifold_iter 10 --gen_descriptor gen_sub3k_iter10

# Model trained with 10k points
python --config results/ABC_10k_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --dataset_name DATASET_NAME --dataset_root data/DATASET_DIR --gen_resolution_global 256

python --config results/ABC_10k_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --dataset_name DATASET_NAME --dataset_root data/DATASET_DIR --gen_resolution_global 256 --manifold_points -1 --noise 0.0 --gen_subsample_manifold 10000 --gen_subsample_manifold_iter 10 --gen_descriptor gen_sub3k_iter10

For faster generation, one would want to use a lower marching cubes parameter --gen_resolution_global, e.g. 128.

DATASET_NAME should be replaced by ABCTest, ABCTestNoiseFree, ABCTestExtraNoise, RealWorld, FamousTest, FamousTestNoiseFree, FamousTestExtraNoisy, FamousTestSparse, FamousTestDense, Thingi10kTest, Thingi10kTestNoiseFree, Thingi10kTestExtraNoisy, Thingi10kTestSparse, Thingi10kTestDense

DATASET_DIR should be replaced by 3d_shapes_abc, 3d_shapes_famous, 3d_shapes_real_world, 3d_shapes_thingi10k


python --config results/ShapeNet_Normals_FKAConv_InterpAttentionKHeadsNet_None/config.yaml --gen_autoscale True --gen_autoscale_target 0.01 --gen_resolution_metric 0.01 --dataset_name DATASET_NAME --dataset_root data/DATASET_NAME --manifold_points -1

DATASET_NAME should be replaced by SceneNet20, SceneNet100, SceneNet500, SceneNet1000.

The --gen_autoscale_target parameter is the average distance between a point and its nearest neighbor in the training set, 0.01 for ShapeNet train set with 3000 points.



python --gendir results/ShapeNet_None_FKAConv_InterpAttentionKHeadsNet_None/gen_ShapeNet_test_3000/ --meshdir meshes --dataset ShapeNet --split test --gtdir data/ShapeNet


python --gendir results/SyntheticRooms_None_FKAConv_InterpAttentionKHeadsNet_None/gen_SyntheticRooms_test_10000/ --meshdir meshes --dataset SyntheticRooms --split test --gtdir data/synthetic_room_dataset 


python --gendir results/ShapeNet_Normals_FKAConv_InterpAttentionKHeadsNet_None/gen_SceneNet20_test_allPts/ --meshdir meshes --gtdir data/SceneNet

ABC, Famous and Thingi10k

For evaluation on the ABCTest please run:

python --gendir results/ABC_3k_FKAConv_InterpAttentionKHeadsNet_None/gen_ABCTest_test_3000 --meshdir meshes/04_pts/ --gtdir data/3d_shapes_abc/abc/

You can adapt the paths to evaluate on the different datasets and noise variants.

Pretrained models

We provide pre-trained models for FKAConv backbone.

ShapeNet 3k, noise, no normals

ShapeNet 3k, no noise, normals

SyntheticRooms 10k

ABC 3k

ABC 10k