panodepth's Introduction

PanoDepth

Getting Started

Requirements

Python (tested on 3.7.4)
PyTorch (tested on 1.4.0)
Other dependencies

Datasets

We train and evaluate on Stanford2D3D, 360D, and 360 stereo dataset.

Usage

Train one-stage monocular depth estimation, run:

python main_mono.py

Train two-stage with the first coarse stage fixed, run:

python main_fullpipeline_pretrain.py

Train two-stage end-to-end, run:

python main_fullpipeline.py

Train 360 stereo matching only, if sample on disparity, use:

python main_stereo_disp.py

Train 360 stereo matching only, if sample on depth, use:

python main_stereo_depth.py

You can change the two-stage configurations in the args.

`--baseline` defines the novel view synthesis baseline from the input view. `--nlabels` defines the number of hypothesis planes for cascade levels. `--interval` defines the depth interval for the second cascade level.

Here are some result comparisons.

If you find our code/models useful, please consider citing our paper:

@inproceedings{li2021panodepth,
  title={PanoDepth: A Two-Stage Approach for Monocular Omnidirectional Depth Estimation},
  author={Li, Yuyan and Yan, Zhixin and Duan, Ye and Ren, Liu},
  booktitle={2021 International Conference on 3D Vision (3DV)},
  pages={648--658},
  year={2021},
  organization={IEEE}
}

panodepth's People

Contributors

Stargazers

Watchers

panodepth's Issues

Request for Trained Model Weights

Hello Yuyan,

You have done a great work. Thanks for the repo. I wanted to run the code with pretrained weights of ResNet360 and psmnet_spherical_up_down_inv_depth. Can you share the checkpoint file that you used for evaluating the full_pipeline (as shown in the attached snapshot)? Thanks.

Recommend Projects