the problem during training about loftr HOT 11 CLOSED

Thank you for your interest in our work. I can't position the problem without further information. On which dataset are you training? Which matcher are you using? When did you meet this problem during the entire training phase? Did it happen from the very start? A training sample with an expec_f_gt value bigger than 1.0 means an invalid sample for the fine-level window used, which would be filtered during the loss calculation. However, we always take parts of the fine-level training samples from the gt coarse correspondences, aiming at alleviating this problem.

from loftr.

I am training in Megadepth dataset. the matcher is dual softmax. this problem happen from the start. I don't sure it can be improved after several epochs. and the coarse loss weight and fine loss weight are all 1. Does this case make training unstable?

from loftr.

Does the training converge with an increasing pose AUC or the training never lead to the pose AUC boosting? At which image resolution are you training? What pose AUC could you get?

from loftr.

I don't get the AUC result now. but during training, the logging always appear assign a false supervision to avoid ddp deadlock. the image resolutiong is 640

from loftr.

What batch_size are you using? I could try to reproduce your problem and give you more feedback then.

from loftr.

from loftr.

hello. Maybe warp_kpts function has some problem.I vis the keypoint0 and w_keypoint0.,but the match is wrong. But I don't sure. whether you are in the same case.

from loftr.

Hi, I tried to train LoFTR in your set-up with the released code, and I could not reproduce your problem. Note that I update the set-up of the MegaDepth depth map here, please make sure you have set up the dataset correctly.

from loftr.

I use d2-net to preprocess the medadepth dataset. Maybe this make camera poses different. Could you share the data preprocess code for me?

from loftr.

Hi, I forgot to mention that we use the d2-net preprocessed images and camera poses as well during training (I will update the doc asap). In addition, the original depth maps from MegaDepth are used.

from loftr.

thank you very much!!

from loftr.