Homography Imitation Learning

This repository holds the source code for:

• Deep Homography Prediction for Endoscopic Camera Motion Imitation Learning
• Deep Homography Estimation in Dynamic Surgical Scenes for Laparoscopic Camera Motion Extraction

Inference code for these papers can be found here.

This repository is under active development and might change. See details below.

Create Environment

Create environment

conda create -n hil_torch2 # names in env_*yml

Install dependencies with mamba (quicker)

mamba env update -f env_hil_torch2.yml

Code Structure

This code is built with PyTorch Lightning. A refactoring is necessary at some point.

• Neural networks under lightning_modules
• Data loaders under lightning_data_modules
• Code is configured via yaml files in config
• Code is run via main_*.py (this needs fix, single main desired)
• Dependencies are handled via Anaconda in env_*.yml files

Papers

Deep Homography Prediction for Endoscopic Camera Motion Imitation Learning

Main idea: Learn camera motion (expressed via homographies) from estimated camera motion.

Source lives under lightning_modules/homography_imitation.

Figure 1 paper: Training pipeline, refer to Section 2.3. From left to right: Image sequences are importance sampled from the video database and random augmentations are applied per sequence online. The lower branch estimates camera motion between subsequent frames, which is taken as pseudo-ground-truth for the upper branch, which learns to predict camera motion on a preview horizon

Citation

@inproceedings{huber2023deeppre,
    author="Huber, Martin and Ourselin, S{\'e}bastien and Bergeles, Christos and Vercauteren, Tom",
    title="Deep Homography Prediction for Endoscopic Camera Motion Imitation Learning",
    booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023",
    year="2023",
    publisher="Springer Nature Switzerland",
    address="Cham",
    pages="217--226",
    isbn="978-3-031-43996-4"
}

Deep Homography Estimation in Dynamic Surgical Scenes for Laparoscopic Camera Motion Extraction

Main idea: Learn camera motion and ignore other motion, such as object motion.

Source lives under lightning_modules/homography_regression

Figure 3 paper: Deep homography estimation training pipeline. Image pairs are sampled from the HFR da Vinci surgery dataset. The homography generation algorithm then adds synthetic camera motion to the augmented images, which is regressed through a backbone DNN.

Citation

@article{huber2022deepest,
  title={Deep homography estimation in dynamic surgical scenes for laparoscopic camera motion extraction},
  author={Huber, Martin and Ourselin, S{\'e}bastien and Bergeles, Christos and Vercauteren, Tom},
  journal={Computer Methods in Biomechanics and Biomedical Engineering: Imaging \& Visualization},
  volume={10},
  number={3},
  pages={321--329},
  year={2022},
  publisher={Taylor \& Francis}
}