search

ufukefe / DFM

Python (Pytorch) and Matlab (MatConvNet) implementations of CVPR 2021 Image Matching Workshop paper DFM: A Performance Baseline for Deep Feature Matching
Other
195 stars 22 forks source link

infrared and visible images pixel level registration #3

Closed ysingfun closed 3 years ago

ysingfun commented 3 years ago

Hi,ufukefe! Thank you so much for your beautiful work! I want to use your algorithm for multi-modal image matching (infrared and visible images).May I ask what modifications I need to make. Thanks!

ufukefe commented 3 years ago

Hi, Thank you for your interest :)

Based on my observations, for multi-modal image matching, you should use the following parameters;

enable_two_stage = False, model = 'VGG19', bidirectional = True, ratio_th = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0] (These parameters mean simply removing the ratio_thresholds (since IR-RGB matching is harder than IR-IR or RGB-RGB marching, we should loose the matches) and using only Stage-1 since (most probably) images are not planar unlike HPatches, if your images are planar then you can arrange enable_two_stage as True)

Also, again based on my observations SuperPoint + SuperGlue, Patch2Pix and LoFTR algoritms work well for multi-modal image matching.