search

vislearn / dsacstar

DSAC* for Visual Camera Re-Localization (RGB or RGB-D)
BSD 3-Clause "New" or "Revised" License
235 stars 36 forks source link

Test with a different camera #31

Open AzizaZhanabatyrova opened 6 days ago

AzizaZhanabatyrova commented 6 days ago

Hi. I trained a model with images from one camera and the performance is good on the test set from the same camera. However when I test the performance of the same trained model with images from a different camera, the poses are a total mess. Is it an expected behavior or could there be any other problem?

ebrach commented 6 days ago

Hi! In general, DSAC can generalise to different camera models if you set the intrinsics properly. However, if the two cameras produce visually very different images, DSAC might fail. If that is the case for you, you might want to think about proper data augmentation for training that closes the domain gap to the test images.