Questions on training about loftr HOT 9 CLOSED

and what is EPI_ERR_THR in default.py ?

from loftr.

1 & 2. Training of the coarse-level LoFTR should converge pretty fast, so I guess there are bugs in the re-implementation, such as incorrect setup of gt coarse matches or loss calculation. Maybe Try overfitting one training sample before moving on.
3. We clip gradients with a norm above 0.5, but I think this should be irrelevant to the problem you met.
4. EPI_ERR_THR is the correctness threshold for a match w.r.t. the squared symmetric epipolar error.

from loftr.

Thank you for your quick response!
I have visualize the gt coarse matches and they are all right.
I am confused about "pad coarse-level matches by randomly sampling from ground-truth matches", is this sampling on the coarse matching? or sample when calculating loss?

from loftr.

1 & 2. Training of the coarse-level LoFTR should converge pretty fast, so I guess there are bugs in the re-implementation, such as incorrect setup of gt coarse matches or loss calculation. Maybe Try overfitting one training sample before moving on.
3. We clip gradients with a norm above 0.5, but I think this should be irrelevant to the problem you met.
4. EPI_ERR_THR is the correctness threshold for a match w.r.t. the squared symmetric epipolar error.

I pad coarse-level matches by randomly sampling from ground-truth matches after coarse mathcing, and as times go, there still is 0 match from coarse mathcing layer. All coarse matches are from gt matches. The coarse loss couldn't convergence. Could you give me more details about training code?

from loftr.

1 & 2. Training of the coarse-level LoFTR should converge pretty fast, so I guess there are bugs in the re-implementation, such as incorrect setup of gt coarse matches or loss calculation. Maybe Try overfitting one training sample before moving on.
3. We clip gradients with a norm above 0.5, but I think this should be irrelevant to the problem you met.
4. EPI_ERR_THR is the correctness threshold for a match w.r.t. the squared symmetric epipolar error.

I found that code 'mconf = conf_matrix[b_ids. i_ids, j_ids]', mconf don't have grad (mconf.requires_grad is False), conf_matrix.requires_grad is True.

from loftr.

The line mconf = conf_matrix[b_ids, i_ids, j_ids] extracts coarse-level matching confidences from the dense confidence matrix. We calculate the coarse loss based on the dense coarse-level matrix instead of the indexing results.

from loftr.

The line mconf = conf_matrix[b_ids, i_ids, j_ids] extracts coarse-level matching confidences from the dense confidence matrix. We calculate the coarse loss based on the dense coarse-level matrix instead of the indexing results.

Thank you very much! I find the line '@torch.no_grad()' which causes the bugs!
And one more question, should I set the 'var.detach()' ?

from loftr.

Sorry, I don't quite get your question. If you are referring to whether to detach the indices for extraction of fine-level patches, then yes. However, in our case, it should already be decorated with torch.no_grad().

from loftr.

Sorry, I don't quite get your question. If you are referring to whether to detach the indices for extraction of fine-level patches, then yes. However, in our case, it should already be decorated with torch.no_grad().

Thank you.

from loftr.