search

facebookresearch / OrienterNet

Source Code for Paper "OrienterNet Visual Localization in 2D Public Maps with Neural Matching"
Other
463 stars 48 forks source link

Generalize to other city is not good #24

Open Bing1002 opened 1 year ago

Bing1002 commented 1 year ago

Hi Sarlin,

Thanks for the fantastic work. I tried several images from different cities, such as Ann Arbor, Mi; Columbus, Ohio; Atlanta, GA. The location is not as I expect. How should this work generalize better to other cities? Thank you!

sarlinpe commented 1 year ago

OpenStreetMap data is pretty similar across large cities, especially within the US. Failure cases can be due to:

  1. incorrect map data (missing labels or wrong location)
  2. lack of distinctive semantic features
  3. large domain gap in terms of image/camera (OrienterNet requires pinhole images and was trained for typical street-level viewpoints, not UAV data).

Would you mind sharing some images and the predictions that you obtain? Thanks.

Bing1002 commented 1 year ago

Hi Sarlin, Thanks for your reply. I just uploaded a few images i tried and they are all performing not good. You can also try those images. I am always wondering about the generalization ability of the model. https://drive.google.com/drive/folders/1CKnVTLbLCS0JPuL7ODjc7AAaidd9aPHz?usp=sharing