amaralibey / gsv-cities Goto Github PK

The paper says that the dataset and code are available. But I don't see the dataset link anywhere on this repo. Am I missing something?

Hi amar,

Thank you for the amazing repo.

I am working on a similar project involving VPR , your data introduced in paper fit me well.

I just wondering when will the full release the code and dataset.

Thx again.

In my device，this result in Nordland is like not good ,I think maybe is parameters is not correct,but I don't know how to adjust which parameters .In addition,I want to ask for you about the groundtruth of dataset how to make.Thank you!

Hi I did not find a license file in your repo. Can you add one please? is it MIT license?
Fantastic work!

Hi? May I ask how to select the query images and to create the following npy file? using MSLS dataset as an example.

self.dbImages = np.load(GT_ROOT+'msls_val/msls_val_dbImages.npy')
self.qIdx = np.load(GT_ROOT+'msls_val/msls_val_qIdx.npy')
self.qImages = np.load(GT_ROOT+'msls_val/msls_val_qImages.npy')
self.ground_truth = np.load(GT_ROOT+'msls_val/msls_val_pIdx.npy', allow_pickle=True)

I am trying to build my own datasets and test your models

Hi,@amaralibey
SPED and Nordland dataset codes are not available. Can you release two dataset code?

To the best of my knowledge, it is not mentioned in your paper. I cannot find a sentence explaining "GSV(Gxxxxx Sxxxxx Vxxxxx)."
It is really confusing for someone who wants to dive into the field of visual place recognition to use short abbreviations without explanations.
Google the word "GSV" would return an even more confusing result.

I would appreciate it if you could help me understand the name.

Hi? May I ask how the gt(pitts30k_test_gt.npy) is arranged? It doesn't seem to be sorted by distance.

Hello, since I am a novice, I have some questions about training. I made an error when running the result main.py file: Missing Logger folder: LOGS/RESNET50/lightning, do I need to change the default path? How? Can I use the MLP feature aggregation method for image retrieval using other datasets?

May I ask how you generate the image "panoid" ? such as "7P0FnGV3k4Fmtw66b8_-Gg"

Hi, Amar
Thanks for the great work.
I just noticed that the implementation of the NetVLAD aggregator is missing in the current code.
I would like to ask how to use GSV-Cities to train NetVLAD.
Would you please release the code for it?

Hi! Thank you so much for your work! I am very new to the space of VPR and would like to ask how a VPR model can be deployed. I have some experience with YOLO computer vision where there is a script to run predictions. May I ask how I can do the same for VPR?

Thank you so much for your help!

Hi amar,

Thank you for the amazing repo.

I'm interested in VPR, however, restricted by machine(GPU), it's a little hard for me to train all the data. So I want to know if you will release the pre-training model.

Thanks

Hi, Amar
Thanks for the great work.
I just noticed that the implementation of the ConvAP module is missing in the current code.
Would you please release the code for it?

Hi There,

Thank you for this excellent code base.

I wanted to ask about batch sizes. In the paper you report results on a batch of 100*4 = 400.

I can't achieve that due to GPU Memory. I am using a batch size of 32 but cannot get my models to converge. Did you experiment with smaller batch sizes ?

Hi,

Thanks for the awesome work! I am new here. There are some problems during reading your code.

What's the vars about GT_ROOT? In your default setting, GT_ROOT is "/home/USER/work/gsv-cities/datasets/"

But how to construct this directory? I am confused.

E.g., I downloaded ESSEX3IN1 dataset. But there are no ESSEX_dbImages.npy, ESSEX_qImages.npy and ESSEX/ESSEX_gt.npy. What should I do?

Hi, Amar
Thanks for the great work.
I found a bug. In main.py, def validation_ epoch(),
the length of feats is far less than num_ References results in utils.get_ validation_ Recall() divisor is 0 and appearing NAN

Hello @amaralibey.
When constructing the training set, the order of places within a city is shuffled, whereas there is an order between cities because buildings within a city share a similar architectural style, whereas architectural styles differ between cities.

Hello，I have some question about training. I met a problem When I tried to run the main.py . According to the prompt, I tried modifying the value of the max_steps, but it still shows max_steps=1. Can you give me some suggestions about this problem?

Hello, @amaralibey ,first of all,many thanks for your work!I really appreciate for your great work :) When I reproduced, I had some problems and needed help.

1、How to test the pitts250k val/test dataset with the trained weights?
2、If something unexpected happens that stops the training, how to continue training on the basis of the existing trained weight?
3、If I want to obtain a list of the top N images retrieved for each query in the test dataset in the test stage.It just need to output a txt file( including the prediction results ).How to to achieve this？
4、I noticed that PCA algorithm was used in the paper, but I didn't find it in the code. May I ask how to implement it？

I am always looking forward to your kind response:)
Best regards.

Hi, Amar
Thanks for the great work.
I just noticed that /datasets/msls_val/msls_val_dbImages.npy、msls_val_qImages.npy、msls_val_pIdx.npy is missing in the current code.
Would you please release the code for it?