search

mode-str / crossview

This repository contains the dataset link and the code for our paper MCCG: A ConvNeXt-based Multiple-Classifier Method for Cross-view Geo-localization.
11 stars 1 forks source link

Request about releasing SUE training code #5

Open fedeloper opened 2 months ago

fedeloper commented 2 months ago

Hi, Can you please also share the training code for SUE Dataset?

mode-str commented 2 months ago

The code for the train and test of the SUES-200 dataset and the code for the University-1652 dataset are the same except for the different file paths they read from. The method in this paper handles the SUES-200 dataset by processing the UAV images taken at different altitudes separately, which means training and testing are done for data at four different heights. It is recommended that you adjust the organization of the images at each altitude of the SUES-200 dataset according to the structure of the University-1652 dataset and the README document.

fedeloper commented 2 months ago

Thank you for your response. I already adhered to the instructions in the README document, ensuring that the structure matches that of SUE exactly. When I configure the path to "train," I encounter a FileNotFoundError: [Errno 2] No such file or directory: './data/sue/Training/satellite.' Conversely, setting the path to an altitude, for example, "150," results in a FileNotFoundError: [Errno 2] No such file or directory: '/home/fred/CrossViewBNN/crossview/data/sue/Training/150/street.' This error is expected since SUE does not contain a 'street' folder. Indeed in make_dataloader the access to this folder is hard coded and there is no differentiation between university or SUE datasets. Are you sure you loaded the full code for training SUE?

mode-str commented 2 months ago

OK, I got it. You may need to change the code about "dataloader" to avoid to load the 'street' and 'google' images. If you do not want to change the code, it seems that it is also possible to add corresponding 'street' images (although these pictures are not used for training) in SUES.

fedeloper commented 2 months ago

Hi thank you, I modified the dataloader to not load these folders and managed to do the training. But the results are worse than the ones on the papers: for 150 m I got: Recall@1:87.50 Recall@5:95.00 Recall@10:97.50 Recall@top1:98.75 AP:83.04 Recall@1:76.70 Recall@5:93.93 Recall@10:97.85 Recall@top1:90.97 AP:80.67

Are you sure all the hyper-parameters are the same? Are there any other modifications to reproduce the results on the paper?

mode-str commented 2 months ago

According to the previous works on the SUES-200, we resize the input images to 384, and we train the model at four different heights respectively.