search

wangch-g / lanet

MIT License
30 stars 12 forks source link

experience on the hyperparameter of the loss #7

Closed cxmscb closed 1 year ago

cxmscb commented 1 year ago

Hi, I retrained the lanet model on one new dataset, and found that the numbers of the detected points is small. The models does not detected the points when the images are corrupted. Could I change the hyperparameter for improving the numbers of the detected points?

cxmscb commented 1 year ago

I read the following code in the loss function. Could it hurt the performance? if n_feat < 20: continue

wangch-g commented 1 year ago

I suggest you finetuning the pre-trained model on the new dataset. Could you show some images in the dataset?

cxmscb commented 1 year ago

Thanks. I found these two images hard to match, and I have no idea to deal with it. img1 img2

cxmscb commented 1 year ago

I change the weight of the consistent-constrain loss, which can help the repeatness.

repeatable_constrain + 1.5 * consistent_constrain + aware_consistent_loss