A geometry-aware deep network for depth estimation in monocular endoscopy

[ Paper ] [ Project ]

Installation

$ git clone https://github.com/YYM-SIA/LINGMI-MR  
$ cd LINGMI-MR
$ pip install -r requirements.txt  

Datasets

You can download EndoSLAM dataset from here and ColonoscopyDepth dataset from here.
You can also download our scenario dataset from here.

Training

Training with our scenario dataset

train_depth.py --config configs/blender_train.json

Test

Run the inference using scenario model

eval_depth.py --config configs/blender_eval.json

Models

Model	Base Network	Abs.Rel.	Sqr.Rel	RMSE	RMSElog	a1	a2	a3
Scenario	ResNet18	0.276	0.017	0.066	0.349	0.517	0.819	0.941

Demo

Citation

If you find our work useful please consider citing our paper:

@article{YANG2023105989,
title = {A geometry-aware deep network for depth estimation in monocular endoscopy},
author = {Yongming Yang and Shuwei Shao and Tao Yang and Peng Wang and Zhuo Yang and Chengdong Wu and Hao Liu},
journal = {Engineering Applications of Artificial Intelligence},
volume = {122},
pages = {105989},
year = {2023},
}

Acknowledgements

Thanks to Shuwei Shao for his excellent work AF-SfMLearner, and Jin Han Lee for his BTS, Ozyoruk for his EndoSLAM, Recasens for his Endo-Depth-and-Motion, Godard for his Monodepth2.