jeff-zilence/vigor

Official repository for VIGOR : Cross-View Image Geo-localization beyond One-to-one Retrieval

vigor's Issues

random choice in dataloader.get_next_batch()

hello, why random choice according to sat instead of directly choice from panorama list? are there any special rules here?

and why only feed the first satellite(aerial) image of each panorama(ground) instead of all of them? (self.train_label[img_idx][0])

THE consult of VIGOR

It is an honor to use the dataset you provided。However,Why did I fill out your questionnaire but did not receive the dataset download method? Could you tell me the specific download method of the dataset?

Looking forward to your reply!!

Same and Cross

I have a question. In the same mode, the street view training pictures are 52,609 and the satellite pictures are 40,007. In the cross mode, the street view training pictures are 51,520 and the satellite pictures are 27,101?

How to get the dataset

Hi,
I have submitted the questionnaire but where I can get the dataset.

Thanks!

question about mining in dataloader.py

hello, line 431 of dataloader.py need self.mining_pool_size >= self.train_sat_data_size, so how to random.sample() a list of n elements from a seq shorter than n? Is this a typo?

data request

Hi,
I have submitted the questionnaire but where I can get the dataset.

Thanks!

camera pose of satellite images

Hi Sijie,

I want to know whether you have the camera pose of satellite images? or you get only original aerial images from google maps api but any other meta information?

In other words, I want to know image-forming principle for your datasets

Best wishes.

Inconsistent GT location between the positive and semi-positive satellite images

I have downloaded the VIGOR dataset (link obtained via authors) and created a Jupyter Notebook to visualize the data. Each street image has 1 positive satellite image and 3 semi-positive satellite images (details about positive and semi-positives are in the VIGOR paper). I see inconsistency between the ground truth location of different satellite images for Seattle, while I do not see this for New York, for example. Therefore, I suspect that there is something wrong with Seattle's labels (and possibly those of other cities as well).

Below is a screenshot from my Jupyter Notebook for street image --Bm25YXcr-aJ-Nbdz0oIw,47.593046,-122.300124,.jpg of Seattle. It can be seen that the yellow dot (the ground truth location) is on the right side of the road for the positive and semi-positive 2 satellite image, while for semi-positives 1 and 3 it is on the left side of the road.

Below is a screenshot from my Jupyter Notebook for street image -DO-hyYo-ICf5irvAKwMMw,40.717596,-73.975717,.jpg of New York. It can be seen that the location here is consistent between the different satellite images.

I would like to know if others can also reproduce this problem and how this possible problem could be solved.

How to evaulate your model?

Hi,

I want to know how to evaulate models without training (only use the checkpoints to evaluate retrieval accuracy)?

Best wishes.

Question about hyperparameters.

Hi,

We want to know the details about the hyperparameter "batch_hard_count" for reproducing the baselines of "SAFA+mining". We also hope you can provide more details about how to set corresponding settings of your 4 results in README.md .

Best

Orientation information

Thanks for the great dataset!

I am working on a project where I require the orientation of the ground vehicle in addition to the panorama image. Since the panoramas are north-aligned, have you considered additionally publishing the vehicle's bearing per frame?

sha256 of the VIGOR dataset

Hi, thanks for your excellent work and for providing the relevant dataset. And I noticed that size of the dataset is about 104GB in total, which means that it is quite a large dataset. However, my computer is relatively not that high-performing, which led me to notice a sub-file download error after almost half an hour of unpacking. While my re-download has resolved the issue, it would be helpful if you could agree that posting the SHA256 checksum of the file. It could help someone with the same issue locate and resolve the problem more quickly. Or, If you allow me to, can I post the SHA256 of the corresponding file below this Issue? :)