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

The performance of ResNet18-dfm It's not as good as VGG19-dfm #12

Open zhongqiu1245 opened 6 months ago

zhongqiu1245 commented 6 months ago

Hello, thank you for your amazing job! I tried to use ResNet18 as the backbone of DFM, but the pairs of keypoints were much less than VGG-DFM(only half, for example, 1267vs2465). Theoretically, the performance of ResNet is much better than VGG I picked up the featuremaps of ResNet18 of original img, conv1_x, conv_2x, conv-3x， conv_4x, conv_5x.(all after ReLU) I noticed that the feature maps which you picked in VGG19 was not following the Stage strictly. Some feature maps were after Conv, some were after ReLU. Could you tell me the rules of picking up feature maps? Than you in advance!

ufukefe commented 6 months ago

Hi, you can check the following paper

zhongqiu1245 commented 6 months ago

@ufukefe Thank you for your reply. Could you release your GCM code in github? Your link (http://mias.group/GCM/) which in paperis not working. It always connects http://mias.group/GCM even you press ['The code can be found in this repository.'] Thank you!