OffsetOPT: Explicit Surface Reconstruction without Normals [arXiv]

OffsetOPT is a method for explicit surface reconstruction from 3D point clouds without relying on point normals. It introduces a two-stage framework: (1) a neural network trained to predict local surface triangles from clean, uniformly sampled point clouds, and (2) a per-point offset optimization applied to new input point clouds — including both small-scale shapes and large-scale scenes — for surface reconstruction. The method is effective for accurate surface reconstruction and sharp detail preservation.

---

🔧 Installation

• All experiments in this project were conducted using the following configuration:

  • GPU: NVIDIA RTX 4090
  • CUDA: 12.1
  • PyTorch: 2.3.0

• We recommend using mamba for fast and reliable environment setup:

  # Install mamba, create the environment, and activate it
  conda install -y mamba -n base -c conda-forge
  mamba env create -f environment.yml -y
  conda activate OffsetOPT 

• All required data can be obtained by running bash download_data.sh. It will download and store all data in the subfolder ./Data of the project directory.

Stage 1: Train the triangulation network

• We provide pretrained models:
  • knn=50: trained_models/model_knn50.pth
  • knn=100: download from [this link] model_knn100.pth and store it in trained_models/
• To train the model from your own side, run:

  python S1_train.py  # the default setting is knn=50

Stage 2: Optimize point offsets to reconstruct surfaces

By default, the provided model_knn50.pth is used.

• Run the following command for point clouds formed by GT mesh vertices:

  python S2_reconstruct.py --delta=0.0 --dataset=ABC_test  # replace 'ABC_test' with `FAUST`, or `MGN` for the respective datasets

• Run the following command for dense point clouds:

  python S2_reconstruct.py --delta=0.02 --dataset=ScanNet
  python S2_reconstruct.py --delta=0.1 --dataset=Matterport3D
  python S2_reconstruct.py --delta=0.1 --dataset=CARLA_1M
  python S2_reconstruct.py --delta=0.01 --dataset=Thingi10k --rescale_delta
  python S2_reconstruct.py --delta=0.015 --dataset=Stanford3D --rescale_delta

  • For Matterport3D, there are 6 scenes (listed in MP3D_5cm.lst) in relatively small scale (<15 meters).
    We reconstruct them using a finer resolution (--delta=0.05, i.e., 5 cm) rather than 10 cm in the released results.

  • The reconstructed meshes of Stanford3D scan data [link] contain bumpy vertices, which can be smoothed using one iteration of Laplacian smoothing using python smooth_Stanford3D.py. The results with and without smoothing are stored in smoothed_Stanford3D/ and Stanford3D/, respectively.

Evaluation of Reconstructed Surfaces

You can download our reconstructed meshes from using bash download_results.sh.

• shape evaluation (sample 100K points):

  # ABC: 6771 meshes — this will take time
  python main_eval_acc.py --gt_path=./Data/ABC/test --pred_path=./results/ABC_test
  
  # FAUST:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/FAUST --pred_path=./results/FAUST
  
  # MGN:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/MGN --pred_path=./results/MGN
  
  # Thingi10K:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/Thingi10K --pred_path=./results/Thingi10K
  
  # Stanford3D:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/Stanford3D --pred_path=./results/Stanford3D
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/Stanford3D --pred_path=./results/smoothed_Stanford3D

• scene evaluation (sample 1 Million points):

  # ScanNet:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/ScanNet --pred_path=./results/ScanNet --sample_num=1000000 
  
  # Matterport3D:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/Matterport3D --pred_path=./results/Matterport3D --sample_num=1000000 
  
  # CARLA:
  python main_eval_acc.py --gt_path=./Data/GT_Meshes/CARLA --pred_path=./results/CARLA_1M --sample_num=1000000 

Citation

If you use this work, please cite:

@article{lei2025offsetopt,
  title={OffsetOPT: Explicit Surface Reconstruction without Normals},
  author={Lei, Huan},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2025}
}