Skip to content

Scene matching for automatically initializing video-based registration algorithms

License

Notifications You must be signed in to change notification settings

AyushiSinha/AutoInitialization

Repository files navigation

Automatic initialization of endoscope in canonical coordinate frame given endoscopic video frame

This software allows a user to generate data simulating a moving endoscope in a textured nasal cavity (or other anatomy) mesh and to train a neural network to learn the regions in which the endoscope must be located to generate particular images. Please cite this preprint if results from this software are used in a paper: "Towards automatic initialization of registration algorithms using simulated endoscopy images", Ayushi Sinha, Masaru Ishii, Russell H. Taylor, Gregory D. Hager, Austin Reiter. arXiv:1806.10748 (2018). URL: https://arxiv.org/abs/1806.10748.

Dependencies:

  • OpenGL: pip install PyOpenGL PyOpenGL_accelerate
  • TriMesh: pip install trimesh
  • PIL: pip install Pillow
  • Torch: pip install torchvision
  • Scikit-image: pip install scikit-image

Run:

  • Clone this repository
  • Run python viewer.py to render default mesh
    • Mesh and texture image can be modified in opengl_viewer/opengl_viewer.py
    • Move the camera inside the mesh using WQAZSX and arrow keys
    • Key c saves the current view as an image and the current camera pose in a text file, each pair known as a keyframe
  • run_data_collection.sh calls collect_data.py to interpolate between saved keyframes and save renderings at different camera poses
  • view_training_data.py displays the saved images in a given folder allowing the user to inpect the training images
  • scene_classifier.py trains a neural network to learn the region that a camera should lie in to generate different images

About

Scene matching for automatically initializing video-based registration algorithms

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published