search

alexandrosstergiou / SoftPool

[ICCV 2021] Code for approximated exponential maximum pooling
MIT License
288 stars 52 forks source link

sfotpool instead of convolution in YOLO #42

Closed xiaoyi-creator closed 2 years ago

xiaoyi-creator commented 2 years ago

I want to use sfotpool instead of convolution in YOLO to complete image downsampling, but I don’t know whether to use 1D, 2D or 3D. Can you give me some suggestions?

alexandrosstergiou commented 2 years ago

Hi @xiaoyi-creator,

If you are using 2D Convolutions I suppose you could SoftPool2d. Do note that you will be basically just downsampling if you do that, so you might want to still use 2D Convolutions but also maintain the spatial dimension. You could then use SoftPool to spatially downsample them. This way you can still have a learnable part in your downsampling.

There are many options ofc since is up to you and how you would like to introduce SoftPool in your project.