What are the full names of "GSV" "VGA" and "dbfnames.mat"? about netvlad HOT 7 CLOSED

from netvlad.

from netvlad.

Well, I am still confused about the dbfnames.mat, it seams you have made two versions of the matlab matrix struct data that are completely different.

1. According to https://github.com/gmberton/VPR-datasets-downloader/blob/main/format_tokyo247.py
   tokyo247.mat is a dictionary that contains a key named "dbStruct".
   The value of tokyo247.mat["dbStruct"] is an array of 2 elements:

• the first element is an image path ending with ".jpg"
• the second element is the db_utms

2. Reading the dbfnames.mat I can see the following structure:
   
   So the actual structure is different from the above repository.
   And I cannot find a documentation explaining how to handle it.

Hi, I'm currently away from a computer and will answer in more detail later. The code in MATLAB is meant to abstract away the details - there is no need to understand what are underlying names in the mat files, the Db classes do everything for you through the provided functions, e.g. you give it queryID and dbID and it tells you if it's a positive or negative. Different datasets are organized differently so this makes it easy to use in an unified way in the MATLAB code. GSV in the new dataset you mentioned likely stands for Google Street View. Best, Relja
…
On Sat, 27 Apr 2024, 10:42 叶璨铭, @.> wrote: To the best of my knowledge, such words are not mentioned in the NetVLAD paper. However, it is used in the naming of the files in dataset Tokyo 247, emerging from the link that you sent me in the email. image.png (view on web) https://github.com/Relja/netvlad/assets/41530341/1eb93f60-322a-4daf-8062-0ad4f2f72278 I am really confused by the short abbreviations and cannot figure out the most likelihood explanation after google them only getting inconsistent results depicting the terms of other fields. I believe it is important for me to understand what is GSV, because there is a new dataset named after GSV, and the short abbreviation is not explained in their paper too. See also amaralibey/gsv-cities#23 <amaralibey/gsv-cities#23> . What's more, I noticed some codebases (like https://github.com/gmberton/VPR-datasets-downloader) using the Tokyo 247 would mention that I shall download the "tokyo247.mat" before I proceed with training. Is "dbfnames.mat" the same as "tokyo247.mat"? I would appreciate it if you could offer me some guidance. — Reply to this email directly, view it on GitHub <#57>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABB3ZUOWOF2WKFGS5KQWFLY7NQIJAVCNFSM6AAAAABG35NFAGVHI2DSMVQWIX3LMV43ASLTON2WKOZSGI3DMOJRGY4TQOI . You are receiving this because you are subscribed to this thread.Message ID: @.>

from netvlad.

Hi,

There are multiple projects by multiple authors so let's clarify:

1. Tokyo 24/7: Created by one set of people, it has the raw files you posted a screenshot of in the first message

2. NetVLAD: Created by a different set of people, I'm the first author. This project introduced another dataset and evaluated on two different datasets. All datasets were in different formats as they were created by different people, so I processed them into a unified format and saved the metadata in tokyo247.mat files and the like. None of this is explained in the paper, because the paper doesn't care about details of how datasets are stored on disk. It's not explained in the NetVLAD README because the NetVLAD code gives a simple API that unifies all the relevant datasets, abstracting away these tedious differences between datasets. However, it is described in the header of this file https://github.com/Relja/netvlad/blob/master/datasets/dbBase.m . Metadata from (1) is not changed, you can certainly find corresponding files and values in NetVLAD metadata and metadata of (1), but the order of images could be shuffled as it doesn't matter as long as metadata associations are preserved.

3. The VPR datasets python code you refer to: created by yet another set of people with no overlap. I never used this code so I can't comment. Presumably they are loading my files instead of files from (1) as they then don't have to repeat the work I did in unifying dataset formats.

So, the question is what is the actual problem you are having? If you want to understand the format NetVLAD uses to store dataset metadata (derived from original metadata), then read the header of https://github.com/Relja/netvlad/blob/master/datasets/dbBase.m , item 3. Does that answer your question?

from netvlad.

Hi,

There are multiple projects by multiple authors so let's clarify:

1. Tokyo 24/7: Created by one set of people, it has the raw files you posted a screenshot of in the first message
2. NetVLAD: Created by a different set of people, I'm the first author. This project introduced another dataset and evaluated on two different datasets. All datasets were in different formats as they were created by different people, so I processed them into a unified format and saved the metadata in tokyo247.mat files and the like. None of this is explained in the paper, because the paper doesn't care about details of how datasets are stored on disk. It's not explained in the NetVLAD README because the NetVLAD code gives a simple API that unifies all the relevant datasets, abstracting away these tedious differences between datasets. However, it is described in the header of this file https://github.com/Relja/netvlad/blob/master/datasets/dbBase.m . Metadata from (1) is not changed, you can certainly find corresponding files and values in NetVLAD metadata and metadata of (1), but the order of images could be shuffled as it doesn't matter as long as metadata associations are preserved.
3. The VPR datasets python code you refer to: created by yet another set of people with no overlap. I never used this code so I can't comment. Presumably they are loading my files instead of files from (1) as they then don't have to repeat the work I did in unifying dataset formats.

So, the question is what is the actual problem you are having? If you want to understand the format NetVLAD uses to store dataset metadata (derived from original metadata), then read the header of https://github.com/Relja/netvlad/blob/master/datasets/dbBase.m , item 3. Does that answer your question?

Thanks for your detailed clarification. My actual problem is that I don't know where to download tokyo247.mat. According to your explanation, I should run the matlab code https://github.com/Relja/netvlad/blob/master/datasets/dbBase.m to process dbfnames.mat in order to get tokyo247.mat, is that right?

from netvlad.

It's explained in the readme but maybe not very well, sorry about that. You can download it from this page
https://www.di.ens.fr/willow/research/netvlad/ , see "All dataset specifications"

You don't have to use dbBase.m, it's just the common interface for the use of these datasets in my NetVLAD matlab code - it sounds like you won't be using matlab but will access the data from python through that other codebase. I just linked to this file because the header explains how tokyo247.mat is structured and you asked for that, e.g. you have a dbStruct matlab structure with arrays such as dbImageFns which gives the database image filenames etc.

from netvlad.

It's explained in the readme but maybe not very well, sorry about that. You can download it from this page https://www.di.ens.fr/willow/research/netvlad/ , see "All dataset specifications"

You don't have to use dbBase.m, it's just the common interface for the use of these datasets in my NetVLAD matlab code - it sounds like you won't be using matlab but will access the data from python through that other codebase. I just linked to this file because the header explains how tokyo247.mat is structured and you asked for that, e.g. you have a dbStruct matlab structure with arrays such as dbImageFns which gives the database image filenames etc.

Thank you very much. That solves my problem now.

from netvlad.