TruongKhang / DeViLoc

[CVPR2024 Oral] Learning to Produce Semi-dense Correspondences for Visual Localization
Apache License 2.0
91 stars 5 forks source link

Regarding the issue with the file '7scenes_ densevlad_detrieval_top_10' #5

Open GottenZZP opened 3 months ago

GottenZZP commented 3 months ago

Hello, thank you very much for open-sourcing this excellent project. I have a question: how can I apply it to my own dataset? I noticed that when evaluating the 7scenes dataset, it is necessary to download a file named "7scenes_densevlad_retrieval_top_10" from your Google Drive, which contains the top 10 retrieved reference images for each query image. If I want to apply this to my own dataset, do I need to manually select the 10 reference images corresponding to each query image every time? image

TruongKhang commented 3 months ago

Hi @GottenZZP, the reference images are selected based on global feature extraction networka such as NetVLad, and CosPlace. You can refer to HLoc pipeline to see how to use these models. We don't manually select top-k reference images for each query image.

GottenZZP commented 3 months ago

Sure! Thank you for your response. However, I have encountered a new issue now. When I was pretraining your model, I downloaded the 199GB Megadepth dataset as you mentioned. The downloaded dataset structure matches the one described in your README file (as shown in Figure 1). But during training, it threw an error saying it couldn't find the image file at this path: "data/megadepth/Undistorted_SfM/0034/images/17498130_ef46860f8c_o.jpg" (as shown in Figure 2). Could you please let me know where I can download this Undistorted_SfM folder? figure1: image figure2: image