A question about train loss

[cslvjt]

When I trained on my own dataset, I found that the loss function did not converge, have you encountered this problem?

[cslvjt]

I'm using the training settings you recommended and the video has a total of 61 frames。

[cslvjt]

Another question about num_pts. What's the point of the num_pts? Thank you.

[qianqianwang68]

Since we used a hard negative mining strategy that samples points with high errors periodically, it is normal for the loss to increase. But it should be decreasing in each [n20k, (n+1)20k] interval. Because you applied strong smoothing I cannot tell if this is true. I suggest also looking at the intermediate visualizations that are saved to have a better sense.

num_pts is the number of point correspondences that we sample for each pair of images in each training batch.

[cslvjt]

Thank you!