NeuralRecon: Real-Time Coherent 3D Reconstruction from Monocular Video

Project Page | Paper

NeuralRecon: Real-Time Coherent 3D Reconstruction from Monocular Video
Jiaming Sun*, Yiming Xie*, Linghao Chen, Xiaowei Zhou, Hujun Bao
CVPR 2021 (Oral Presentation and Best Paper Candidate)

real-time video

TODO List and ETA

  • Code (with detailed comments) for training and inference, and the data preparation scripts (2021-5-2).
  • Pretrained models on ScanNet (2021-5-2).
  • Real-time reconstruction demo on custom ARKit data with instructions (2021-5-7).
  • Evaluation code and metrics (expected 2021-6-10).

How to Use


# Ubuntu 18.04 and above is recommended.
sudo apt install libsparsehash-dev  # you can try to install sparsehash with conda if you don't have sudo privileges.
conda env create -f environment.yaml
conda activate neucon
[FAQ on environment installation]
  • AttributeError: module 'torchsparse_backend' has no attribute 'hash_forward'

    • Clone torchsparse to a local directory. If you have done that, recompile and install torchsparse after removing the build folder.
  • No sudo privileges to install libsparsehash-dev

    • Install sparsehash in conda (included in environment.yaml) and run export CPLUS_INCLUDE_PATH=$CONDA_PREFIX/include before installing torchsparse.
  • For other problems, you can also refer to the FAQ in torchsparse.

Pretrained Model on ScanNet

Download the pretrained weights and put it under
You can also use gdown to download it in command line:

mkdir checkpoints && cd checkpoints
gdown --id 1zKuWqm9weHSm98SZKld1PbEddgLOQkQV

Real-time Demo on Custom Data with Camera Poses from ARKit.

We provide a real-time demo of NeuralRecon running with self-captured ARKit data.
Please refer to for details.

Data Preperation for ScanNet

Download and extract ScanNet by following the instructions provided at

[Expected directory structure of ScanNet (click to expand)]

You can obtain the train/val/test split information from here.

│   └───scans
│   |   └───scene0000_00
│   |       └───color
│   |       │   │   0.jpg
│   |       │   │   1.jpg
│   |       │   │   ...
│   |       │   ...
│   └───scans_test
│   |   └───scene0707_00
│   |       └───color
│   |       │   │   0.jpg
│   |       │   │   1.jpg
│   |       │   │   ...
│   |       │   ...
|   └───scannetv2_test.txt
|   └───scannetv2_train.txt
|   └───scannetv2_val.txt

Next run the data preparation script which parses the raw data format into the processed pickle format.
This script also generates the ground truth TSDFs using TSDF Fusion.

[Data preparation script]

# Change PATH_TO_SCANNET and OUTPUT_PATH accordingly.
# For the training/val split:
python tools/tsdf_fusion/ --data_path PATH_TO_SCANNET --save_name all_tsdf_9 --window_size 9
# For the test split
python tools/tsdf_fusion/ --test --data_path PATH_TO_SCANNET --save_name all_tsdf_9 --window_size 9

Inference on ScanNet test-set

python --cfg ./config/test.yaml

The reconstructed meshes will be saved to PROJECT_PATH/results.

Evaluation on ScanNet test-set

python tools/ --model ./results/scene_scannet_release_fusion_eval_47 --n_proc 16

Note that uses pyrender to render depth maps from the predicted mesh for 2D evaluation.
If you are using headless rendering you must also set the enviroment variable PYOPENGL_PLATFORM=osmesa
(see pyrender for more details).

You can print the results of a previous evaluation run using

python tools/ --model ./results/scene_scannet_release_fusion_eval_47

Training on ScanNet

Start training by running ./
More info about training (e.g. GPU requirements, convergence time, etc.) to be added soon.


#!/usr/bin/env bash
python -m torch.distributed.launch --nproc_per_node=2 --cfg ./config/train.yaml

The training is seperated to two phases and the switching between phases is controlled manually for now:

  • Phase 1 (the first 0-20 epoch), training single fragments.

  • Phase 2 (the remaining 21-50 epoch), with GRUFusion.


If you find this code useful for your research, please use the following BibTeX entry.

  title={{NeuralRecon}: Real-Time Coherent {3D} Reconstruction from Monocular Video},
  author={Sun, Jiaming and Xie, Yiming and Chen, Linghao and Zhou, Xiaowei and Bao, Hujun},


We would like to specially thank Reviewer 3 for the insightful and constructive comments. We would like to thank Sida Peng , Siyu Zhang and Qi Fang for the proof-reading.
Some of the code in this repo is borrowed from MVSNet_pytorch, thanks Xiaoyang!


This work is affiliated with ZJU-SenseTime Joint Lab of 3D Vision, and its intellectual property belongs to SenseTime Group Ltd.

Copyright SenseTime. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
See the License for the specific language governing permissions and
limitations under the License.


View Github