Matchable Image Transformations for Long-term Visual Localization

Code to accompany our paper "Learning Maximally Matchable Image Transformations for Long-term Metric Visual Localization".

Dependencies

• numpy
• matplotlib
• pytorch 1.1 + torchvision
• Pillow
• progress (for progress bars in train/val/test loops)
• pyviso2 (included)
• tensorboardX

Contents

• run_vkitti.py - Top-level script for training and testing models using the Virtual KITTI dataset.
• run_inthedark.py - Top-level script for training and testing models using the UTIAS In The Dark dataset.
• matchability/ - Main codebase.
  • datasets/ - Modules for loading and pre-processing the Virtual KITTI and UTIAS In The Dark datasets for use with PyTorch.
  • experiment.py - General training, validation and testing loops.
  • models.py - Definitions for matcher proxy and image transformation models.
  • networks.py - General-purpose neural network classes, specialized in models.py.
  • options.py - Defines a container class to store configuration parameters and provides common defaults.
  • transforms.py - Custom image/tensor transformations.
  • utils.py - Miscellaneous utility functions.
• scripts/ - Scripts for producing figures and videos used in the paper.