search

ethz-asl / hfnet

From Coarse to Fine: Robust Hierarchical Localization at Large Scale with HF-Net (https://arxiv.org/abs/1812.03506)
MIT License
779 stars 185 forks source link

Would it be better if i fine-turn the model on my dataset？ #12

Closed lfydegithub closed 5 years ago

lfydegithub commented 5 years ago

Hi, author. thank you for your good job！I have run the hfnet for retrieval and the result is perfectly good. But i wanna know that Would it be better if i fine-turn the model on my dataset？ i have no ground truth, learn your code, i find the GT is from NetVLAD and SuperPoint, so i'm not sure it's useful to fine-train... thanks for your time！

sarlinpe commented 5 years ago

Hi there,

You could indeed run the distillation on your dataset if you believe that it is substantially different from the one we used (Google Landmarks). All is required is a folder of diverse images, ideally in a large number (>200k).

If you only need retrieval and do not care about efficiency, then you could use NetVLAD. If you still want to use HF-Net, then you could first check that it actually performs sufficiently lower than the baseline NetVLAD+SuperPoint on your kind of data. If so, you can do distillation following the instructions here (see also issue https://github.com/ethz-asl/hfnet/issues/5).