No groundtruth coarse match found for

[rasulkhanbayov]

Hello, I am training LoFTR on custom dataset, but in each epoch it gives me warnings like this:

Epoch 0: 83%| | 850/1021 [02:12<00:26, 6.41it/s, loss=3.74, v_num=16]

2025-07-02 17:01:21.930 | WARNING | src.loftr.utils.supervision:spvs_coarse:100 - No groundtruth coarse match found for: [['data/OCTsynthetic/train\19_enface.png'], ['data/OCTsynthetic/train\19_DataCapture.png']]

2025-07-02 17:01:21.976 | WARNING | src.losses.loftr_loss:_compute_fine_loss_l2_std:163 - assign a false supervision to avoid ddp deadlock

in total 187 out 1021 cases it doesn't find gt coarse match.

Does anyone know how can I solve it? Or will it decrease my accuracy, should I definetly solve it?

Thanks in advance!

Best,
Rasul

[QibiaoWu]

Hello, have you solved this problem? If so, could you share your experience? I also encountered the same issue.

[rasulkhanbayov]

No, I haven't unfortunately. It trains with this warning, but gives not good accuracy.

[QibiaoWu]

No, I haven't unfortunately. It trains with this warning, but gives not good accuracy.

This error seems to indicate that the coarse matching did not produce any usable feature points. I'm now questioning whether it's because the input image is too smooth. I used a pair of airway images (the airway walls are usually smooth). During the construction of the dataset, I once tried to use colmap to generate it, but colmap reported an error that it couldn't effectively extract feature points for the data I input. So later, I studied the structure of the dataset by myself and manually created the dataset according to the given data format. It's possible that there was an error during this process.