chengdazhi / Deformable-Convolution-V2-PyTorch

Deformable ConvNets V2 (DCNv2) in PyTorch
MIT License
1.44k stars 229 forks source link

Initialization of offset layer #66

Open bluesky314 opened 4 years ago

bluesky314 commented 4 years ago

According to the paper and implementation the weights in the offset layer are initialised to 0 so the shift in x,y is 0 and modulation is 0.5. Doesnt it make more sense to initialise the modulation to 1 so we can use deformConv as a drop in replacement in any cnn without disturbing its performance.

According to eq 1) in paper it seems like that would make most sense

image