search

Nanne / pytorch-NetVlad

Pytorch implementation of NetVlad including training on Pittsburgh.
433 stars 109 forks source link

Question about the training dataset #59

Closed hyeongilee closed 3 years ago

hyeongilee commented 3 years ago

Hi, Thanks for providing the code. It seems that the distance between features was used to obtain positive and negative data.(in getitem of QueryDatasetFromStruct) Shouldn't the utm(gps) distance be used to obtain positive and negative datasets when training?

Nanne commented 3 years ago

Potential positives and negatives are initially selected based on utm distance: https://github.com/Nanne/pytorch-NetVlad/blob/master/pittsburgh.py#L198-L206

From these potentials the final selections are made according to feature distance, this is in accordance with the Netvlad approach described in the paper.