MDE

Monocular absolute depth estimation from endoscopy via domain-invariant feature learning and latent consistency arxiv link

Core Idea

Because real endoscopic images (typically) do not have ground truth depth, this work leverages synthetically rendered images with known depth with information from real endoscopic images to improve depth estimation performance on real data

Motivation

Qualitative results

Usage

Installation

conda create -n mde python=3.9
conda activate mde
pip install -r requirements.txt # or conda env create -f environment.yml

Train

python train.py --name <your running name> --json_path <your json file path> --min_depth <your minimum depth value> --max_depth <your maximum depth value> 

The code will automatically create a folder to store logs under /src/checkpoints/your running name/

For training the domain alignment version, use train_ours.py. The detailed split information or format can be viewed in create_dataset folders.

Test

python test.py

you need to change the pretrained_path, name and other arguments in the script

If you find this repository useful, please consider citing this paper:

@inproceedings{li2026mde,
  title={Monocular absolute depth estimation from endoscopy via domain-invariant feature learning and latent consistency},
  author = {Li, Hao and Lu, Daiwei and d'Almeida, Jesse and Isik, Dilara and Khodapanah Aghdam, Ehsan and DiSanto, Nick and Acar, Ayberk and Sharma, Susheela and Wu, Jie Ying and Webster III, Robert J. and Oguz, Ipek},
  booktitle={Medical Imaging 2026: Image Processing},
  volume={in press},
  year={2026},
  organization={SPIE}
}