search

Parskatt / DeDoDe

[3DV 2024 Oral] DeDoDe 🎶 Detect, Don't Describe --- Describe, Don't Detect, for Local Feature Matching
MIT License
380 stars 19 forks source link

question about detector and discriptor #34

Open ye450450 opened 3 months ago

ye450450 commented 3 months ago

Is there a reason why discriptor and detector used VGG19 network architecture? Not using ResNet or other network architecture

Parskatt commented 3 months ago

Not really, you might get some improvement by updating the backbone.

Parskatt commented 3 months ago

Vgg19 features tend to be precisely localizable, which is good. See e.g. the ablation in roma. But its usually worse for global matching. You may also want to have a look at more modern backbones not trained on imagenet

ye450450 commented 3 months ago

and i have one more question. In web-ui code, there is https://github.com/Parskatt/DeDoDe/releases/download/dedode_pretrained_models/dedode_descriptor_B.pth urls for getting the descriptor model. But when i get there, there is no file. I cannot find the file. Can i get the descriptor training recipe? and there is descriptor_L file, too?

ye450450 commented 3 months ago

Sorry, there is file. But, there is no directory of https://github.com/Parskatt/DeDoDe/releases.

Parskatt commented 3 months ago

Correct, those are just files. Code for training models can be found in the experiments folder of the repo.