PlanarRecon: Real-time 3D Plane Detection and Reconstruction from Posed Monocular Videos

Project Page | Paper

PlanarRecon: Real-time 3D Plane Detection and Reconstruction from Posed Monocular Videos Yiming Xie, Matheus Gadelha, Fengting Yang, Xiaowei Zhou, Huaizu Jiang CVPR 2022

real-time video

How to Use


conda env create -f environment.yaml
conda activate planarrecon

Follow instructions in torchsparse to install torchsparse.

Pretrained Model on ScanNet

Download the pretrained weights and put it under PROJECT_PATH/checkpoints/release. You can also use gdown to download it in command line:

gdown --id 1XLL5X2M5BPo89An4jom5s0zhQOiyS_h8

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_raw
│   |   └───scene0000_00
│   |       └───scene0000_00.aggregation.json
│   |       └───scene0000_00_vh_clean_2.labels.ply
│   |       └───scene0000_00_vh_clean_2.0.010000.segs.json
│   |       │   ...
|   └───scannetv2_test.txt
|   └───scannetv2_train.txt
|   └───scannetv2_val.txt
|   └───scannetv2-labels.combined.tsv

Next run the data preparation script which parses the raw data format into the processed pickle format. This script also generates the ground truth Planes. The plane generation code is modified from PlaneRCNN.

[Data preparation script]

# Change PATH_TO_SCANNET accordingly.
# For the training/val split:
python tools/ --data_path PATH_TO_SCANNET --save_name planes_9/ --window_size 9 --n_proc 2 --n_gpu 1

Inference on ScanNet val-set

python --cfg ./config/test.yaml

The planes will be saved to PROJECT_PATH/results.

Evaluation on ScanNet val-set

Evaluate 3D geometry:

python tools/ --model ./results/scene_scannet_release_68 --n_proc 16

Training on ScanNet

Start training by running ./


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

Similar to NeuralRecon, the training is seperated to three phases and the switching is controlled manually for now:

  • Phase 1 (the first 0-20 epoch), training single fragments. MODEL.FUSION.FUSION_ON=False, MODEL.TRACKING=False

  • Phase 2 (21-35 epoch), with GRUFusion. MODEL.FUSION.FUSION_ON=True, MODEL.TRACKING=False

  • Phase 3 (the remaining 35-50 epoch), with Matching/Fusion. MODEL.FUSION.FUSION_ON=True, MODEL.TRACKING=True

More info about training to be added soon.

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

There is a tutorial introduced in NeuralRecon. We also provide the example data captured using iPhoneXR. Incrementally saving and visualizing are not enabled in PlanarRecon for now.


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

  title={{PlanarRecon}: Real-Time {3D} Plane Detection and Reconstruction from Posed Monocular Videos},
  author={Xie, Yiming and Gadelha, Matheus and Yang, Fengting and Zhou, Xiaowei and Jiang, Huaizu},


Some of the code and installation guide in this repo is borrowed from NeuralRecon! We also thank Atlas for the 3D geometry evaluation.


View Github