Image-Matching

A demo using SuperGlue and SuperPoint to do the image matching task.

image matching

Overview

• Deeplearning Framework: PyTorch
• use SuperPoint to extract feature points and descriptors from the image
• use SuperGlue to do image matching
• Matching is a class that combine the above two classes
• use Arg class to make configurations

Usage

• use args=Arg() to initialize configurations
• Since the model has two modes to make more concise predicitions, if the scene is indoor, set scene_mode in Arg ‘indoor’. Similarly set the scene_mode ‘outdoor’ if the scene is outdoor
• in class Arg, use set_ method to set parameters
• To be clear, the anchor image should be set first in Arg by using Arg.set_anchor(anchor_name), the image to test (if it can match the anchor image) should be set using the method Arg.set_match(match_name)
• After setting the anchor image, you need not set it again unless you want to make another matching with a different anchor
• use get_similarity(args,img_name) to get the final score
• If the matching score is larger than 0.2, it is likely can be matched
• libtorch-superpoint is a test file, it can successfully use SuperPoint within cpp projects. However, it takes some time to config makefiles. If all goes well, you can cd build -> cmake.. -> make -> ./a and you can get the result. a is an executable file generated by cmake.
• test_transfer is a try to change some parameters so that I can transfer it to ONNX or Libtorch. I changed the dictionary to list and deleted some unuseful code. It may run a little faster than the main file.

To-dos

• Improve the matching speed, now it takes 3s to do a matching