Skip to content

Latest commit

 

History

History
19 lines (14 loc) · 1.32 KB

backbone.md

File metadata and controls

19 lines (14 loc) · 1.32 KB

Backbone

Overview

Adding a backbone should be a fairly straightforward process. The main backbone we use for SiLK is the ParametricVGG found in lib/backbones/superpoint/vgg.py.

A backbone is a PyTorch Module with a few constraints :

  • It should take images as input
  • It should output dense feature maps per image
  • It should subclass CoordinateMappingProvider and implement a mappings() method that provides the bijective coordinate map between the feature coordinate space and the input coordinate space (in order to know where an output feature is located in the image). Convenient functions are provided to automate the process, and one can get inspiration from how ParametricVGG.mappings() is implemented.
  • Once the new backbone is implemented, its configuration file should be added to etc/backbones/.

We can help

We make heavy use of hydra to handle compositionality of our configurations. This can feel a bit daunting at first. So we can help you with the integration of a new backbone, by following the steps below:

  1. Implement the backbone and add the code to lib/backbones, with simple unit tests.
  2. Create a pull request with the title Integrate <name> backbone.
  3. At that point, we can help and iterate on the pull request to make sure the backbone is properly integrated.