A problem with dataset "Tokyo247" about netvlad

[Relja]

Please reread the paper and other relevant papers - we follow a standard procedure used to evaluate location recognition. Tokyo 24/7 is an existing dataset, we didn't define the evaluation procedure for it, read that paper. Specifically: dataset in general have training, validation and test splits. In principle, for EACH split, there are query and database images. Specifically, Tokyo 24/7 is a dataset that is test-only, it doesn't have training or validation, so to train we introduce the Tokyo Time Machine dataset, which has training and validation splits. To evaluate, you take the test dataset (Tokyo 24/7) and search with each of the query images in turn to evaluate the performance.

Definitions of the potential positives and triplet can be found in the paper text:

To create the training tuple for a query, we use all of
its potential positives (images within 10 meters), and we
perform randomized hard negative mining for the negatives
(images further away than 25 meters). The mining is done
by keeping the 10 hardest negatives from a pool of 1000 randomly
sampled negatives and 10 hardest negatives from the
previous epoch. We find that remembering previous hard
negatives adds stability to the training process.