search

irfanICMLL / structure_knowledge_distillation

The official code for the paper 'Structured Knowledge Distillation for Semantic Segmentation'. (CVPR 2019 ORAL) and extension to other tasks.
BSD 2-Clause "Simplified" License
702 stars 104 forks source link

Question about the pair-wise loss #31

Closed colorjam closed 4 years ago

colorjam commented 4 years ago

Hi, thank you for your interesting work. I'm curious about the \alpha in Eq.2. The definition of \alhpa is the connection range, however, in the CriterionPairWiseforWholeFeatAfterPool there is no parameter related to \alpha. Would you mind explain the implementation of \alpha in Eq2.?

irfanICMLL commented 4 years ago

It is a discussion about the pair-wise loss. The results show that keeping a fully connect range is the best choice considering the performance, and use a large \beta will help improve the results. We use a pooling operator for the \beta to achieve the best choice in our paper.

On P6 of our paper, we can get a conclusion from the table. "One can choose to use the local patch to decrease the number of the nodes, instead of decreasing the connection range for a better trade-off between efficiency and accuracy"