search

LiWentomng / boxlevelset

The code for "Box-supervised Instance Segmentation with Level Set Evolution(ECCV2022)"
Apache License 2.0
185 stars 25 forks source link

Question regarding the LevelSetLoss implementation #6

Closed QianyiWu closed 1 year ago

QianyiWu commented 1 year ago

Thanks for your awesome work.

I found in the implementation of LevelSetLoss that the term of length regularization is not calculated. Would you like to share the reason for it?

Besides that, what is the motivation that only calculates the loss for the positive samples in Eqn.(11) in the paper?

Best

LiWentomng commented 1 year ago

@QianyiWu Thanks.

In the ablation study, the length regularization of Level set loss is adopted, which is beneficial for the performance (especially when the performance is low). In the main results, I found that the length regularization can bring the 0.2~0.4mAP drop. So, I commented the calculation for the better final results.

About only calculting the the loss for the positive samples, this follows the original SOLOv2 network. The positive samples comes from the label(target) assignment considering the GT box location and mask prediction. Without this scheme, the network convergence is very slow, and it can't even work.