search

lim-anggun / FgSegNet_v2

FgSegNet_v2: "Learning Multi-scale Features for Foreground Segmentation.” by Long Ang LIM and Hacer YALIM KELES
https://arxiv.org/abs/1808.01477
Other
149 stars 43 forks source link

Design of decoder #7

Closed fsk119 closed 5 years ago

fsk119 commented 6 years ago

Dear author, thanks for your paper about change detection. I' m very interested in the design of decoder in your architecture. However, I am curious why you combine information in that way : alpha*f + f where alpha is average pooling of information from upper layer. Numerous papers tend to implement network like U-net in semantic segmentation field.

lim-anggun commented 5 years ago

Hi, Instead of using previous knowledge from the encoder part to refining object boundaries in the decoder part (like semantic segmentation), we try something difference. We perform global average pooling of 64 features to scale the features in the decoder. This acts like giving importance to each feature slice in the decoder according to pooled weights from the encoder. We found that this way could help improving our network performance in overall (Please refer to the paper for M-FPM vs FPM ablation study). We also provide some visualizations in the journal version of this preprint paper (unavailable just for now).

feel free to reopen this issue if you have further question.