search

vdvchen / SGMNet

Implementation of "Learning to Match Features with Seeded Graph Matching Network" ICCV2021
MIT License
136 stars 22 forks source link

if we change the two image of order, we got the different size of matching-poing, ps : config, extractor:num_kpt=300 #5

Closed larsoncs closed 2 years ago

larsoncs commented 2 years ago

thanks

vdvchen commented 2 years ago

Hi,

Due to the concate operation in our network, the obtained matches are expected to differ when reversing img order. However, statistically the result should make very little differences.

larsoncs commented 2 years ago

can you eliminate the differences?

vdvchen commented 2 years ago

As a workaround, you may run the network twice for both img orders and find the intersection set, as is done for some neural outlier rejector(filtering based methods)