search

AlexYouXin / Explicit-Shape-Priors

35 stars 1 forks source link

About Shape Prior Module #11

Open Billy-ZTB opened 4 days ago

Billy-ZTB commented 4 days ago

Hello! @AlexYouXin Thanks for your great work and contribution! I read your paper and code but still got confused about the SPM you proposed. You generated a learnable shape prior as an input for the SPM, which is expected to bring prior knowledge of the segments' shape to the model. However, the learnable shape prior tensor is generated randomly, neither is it generated by learning the distribution of images and labels in the dataset nor by some sort of shape supervision. I wonder how this randomly initialized parameter achive your goal. I am looking forward to your kind reply!

AlexYouXin commented 4 days ago

Thank you for your attention and interests on this work!! The innitialized parameters serve as the auxiliary input with coarse shape priors, which could cover the whole dataset. During inference, coarse shape priors will be iteratively refined under the interaction with skipped features. Then raw skipped features will be enhanced with more abundant textures, and shape priors will contain more finegrained shape information to further boost the final segmentation as revealed by the qualitative results To some extent, randomly initialized parameters will be updated to the state which is not random after the training process.

Billy-ZTB commented 2 days ago

Thank you for your kind reply! I understand that this learnable parameter can learn some features of the images, but since it's randomly initiated, it can't bring any prior knowledge(I don't know if I am right about this part); further more, there is no specific supervision about the segments' shape, how do you identify that this param serves as shape prior? I am new in this field, this really confuses me.

---- Replied Message ---- | From | @.> | | Date | 09/16/2024 14:44 | | To | @.> | | Cc | Zhang @.>@.> | | Subject | Re: [AlexYouXin/Explicit-Shape-Priors] About Shape Prior Module (Issue #11) |

Thank you for your attention and interests on this work!! The innitialized parameters serve as the auxiliary input with coarse shape priors, which could cover the whole dataset. During inference, coarse shape priors will be iteratively refined under the interaction with skipped features. Then raw skipped features will be enhanced with more abundant textures, and shape priors will contain more finegrained shape information to further boost the final segmentation as revealed by the qualitative results To some extent, randomly initialized parameters will be updated to the state which is not random after the training process.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>