MVSDF – Learning Signed Distance Field for Multi-view Surface Reconstruction


This is the official implementation for the ICCV 2021 paper Learning Signed Distance Field for Multi-view Surface Reconstruction

In this work, we introduce a novel neural surface reconstruction framework that leverages the knowledge of stereo matching and feature consistency to optimize the implicit surface representation. More specifically, we apply a signed distance field (SDF) and a surface light field to represent the scene geometry and appearance respectively. The SDF is directly supervised by geometry from stereo matching, and is refined by optimizing the multi-view feature consistency and the fidelity of rendered images. Our method is able to improve the robustness of geometry estimation and support reconstruction of complex scene topologies. Extensive experiments have been conducted on DTU, EPFL and Tanks and Temples datasets. Compared to previous state-of-the-art methods, our method achieves better mesh reconstruction in wide open scenes without masks as input.

How to Use

Environment Setup

The code is tested in the following environment (manually installed packages only). The newer version of the packages should also be fine.

  - cudatoolkit=10.2.89
  - numpy=1.19.2
  - python=3.8.8
  - pytorch=1.7.1
  - tqdm=4.60.0
  - pip:
    - cvxpy==1.1.12
    - gputil==1.4.0
    - imageio==2.9.0
    - open3d==0.13.0
    - opencv-python==
    - pyhocon==0.3.57
    - scikit-image==0.18.3
    - scikit-learn==0.24.2
    - trimesh==3.9.13
    - pybind11==2.9.0

Data Preparation

Download preprocessed DTU datasets from here


cd code
python training/ --data_dir <DATA_DIR>/scan<SCAN>/imfunc4 --batch_size 8 --nepoch 1800 --expname dtu_<SCAN>

The results will be written in exps/mvsdf_dtu_<SCAN>.

Trained Models

Download trained models and put them in exps folder. This set of models achieve the following results.

Chamfer PSNR
24 0.846 24.67
37 1.894 20.15
40 0.895 25.15
55 0.435 23.19
63 1.067 26.24
65 0.903 26.9
69 0.746 26.54
83 1.241 25.15
97 1.009 25.71
105 1.320 26.48
106 0.867 28.81
110 0.842 23.16
114 0.340 27.51
118 0.472 28.46
122 0.466 27.71
Mean 0.890 25.72


python evaluation/ --data_dir <DATA_DIR>/scan<SCAN>/imfunc4 --expname dtu_<SCAN> [--eval_rendering]

add --eval_rendering flag to generate and evaluate rendered images. The results will be written in evals/mvsdf_dtu_<SCAN>.


cd mesh_cut
python build_ext -i  # compile
python <IN_OBJ> <OUT_OBJ> [--thresh 15 --smooth 10]

Note that this part of code can only be used for research purpose. Please refer to mesh_cut/IBFS/license.txt


Apart from the official implementation, you can also use my re-implemented evaluation script.


If you find our work useful in your research, please kindly cite

	title={Learning Signed Distance Field for Multi-view Surface Reconstruction},
	author={Zhang, Jingyang and Yao, Yao and Quan, Long},
	journal={International Conference on Computer Vision (ICCV)},


View Github