questions about grouth.npz

[WentangChen]

hi thank you for sharing your excellent work! I am trying to make my ground_truth.npz for my own dataset. but I have some questions:

1. does 'utmQ' and 'utmQ' contains the utm coordinates of all pictures in query and database dataset?
2. I have create my datasets with 7 places. Each place have 400 pictures for database and 80 pictures for query. is the number of pictures enough for training (simply using the traning code provided for maliipary dataset)?
3. I have try training on my own dataset (descirbed in question 2.). And I got the following result:

====> Calculating recall @ N
====> Recall@1: 0.0600
====> Recall@5: 0.1200
====> Recall@10: 0.1800
====> Recall@20: 0.3000
====> Recall@50: 0.6000
====> Recall@100: 0.8200

I wonder the result is good or not?

I am looking forward to your reply , thank you!

[Tobias-Fischer]

Hi,

1. Yes
2. No, this is most likely not sufficient. But why not use a pre trained model and use that to extract features on your dataset?
3. The results look poor considering there are only 7 places.

You might want to read our recent “Visual Place Recognition: A Tutorial” paper for some more intuitions.

[WentangChen]

I am sorry for keeping this issue open. I still have some questions:

1. since I wanna adopt your work to train on my on dataset specificlly, Can you give some advice to improve my performance?
2. can you give me some instructions to evaluates the real-time performance of the model?
3. can you tell me how to set my posDistThr and negDistThr value? note that the all 7 places of my datasets are in the same buildings which means the utm of them are quite close.

Looking forward to your reply, thanks!