Skip to content

yuyanli0831/OmniFusion

Repository files navigation

OmniFusion

This repository hosts the original implementation of CVPR 2022 (oral) paper "OmniFusion: 360 Monocular Depth Estimation via Geometry-Aware Fusion ". ArXiv link

Github_gif_prepare2

Citation

If you found our code helpful for your research, please cite our paper as:

@inproceedings{Li2022CVPR,
  Title      = {OmniFusion: 360 Monocular Depth Estimation via Geometry-Aware Fusion},
  Author     = {Li, Yuyan and Guo, Yuliang and Yan, Zhixin and Huang, Xinyu and Ye, Duan and Ren, Liu},
  Booktitle  = {2022 Conference on Computer Vision and Pattern Recognition (CVPR)},
  Year       = {2022},
  Address    = {New Orleans, USA},
  Month      = jun,
}

Pipeline

Github_gif_prepare

Getting Started

Requirements

  • Anaconda (tested on 4.9.2)
  • Python (tested on 3.7.4)
  • PyTorch (tested on 1.9.0)
  • torchvision (tested on 0.10.0)
  • CUDA (tested on 11.6)
  • Other dependencies
git clone https://github.com/yuyanli0831/OmniFusion
cd OmniFusion
python3 -m venv omni-venv
source omni-venv/bin/activate
pip3 install -r requirements.txt

Datasets

We trained and evaluated our method on three datasets Stanford2D3D, 360D, Matterport3D.

FYI, a standard pre-process of the original Matterport3D dataset is applied for the development of 360-image-based depth estimation. Given the downloaded original dataset, the 360 images and depth maps are rendered by a matlab scrpt included in this repo.

Pre-trained models

Our pre-trained models are provided for direct evaluation.

Training

python train_erp_depth.py --fov 80 --patchsize (256, 256) --nrows 4

You can specify the patch fov, patch resolution, patch alignment(3, 4, 5, 6 rows).

Evaluation

python test.py --fov 80 --patchsize (256, 256) --nrows 4

Visualization

For training, visual results will be saved into tensorboard. In addition, depth maps and point clouds will be saved locally during training/evaluation.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published