The number of filter for generate offset is [num_deformable_groups * 2 * kH * kW]?

[gzhcv]

Inside the paper of Deformable ConvNets V2, a modulation scalars was used to modulate the input feature amplitudes from different spatial locations/bins. As the following description,

So, i'm curious about the usage example of DeformConv() in test.py , which setting the number of filter for generate offset as num_deformable_groups * 2 * kH * kW. Shouldn't it be num_deformable_groups * 3 * kH * kW?

Thanks a lot for your code, hope for your reply.

[gzhcv]

One more question, i trained a Deformable ConvNets for Classification task and have gained accurancy. Then i extract the value of feature map used for predicting the offset, the value range from -99 to 99 approximately. This result also bothered me, whether the value should be squeezed to a reasonable range, such as [-5, 5]

[xvjiarui]

num_deformable_groups * 2 * kH * kW is for normal Deformable Conv.
num_deformable_groups * 3 * kH * kW is for modulated Deformable Conv.

[gzhcv]

@xvjiarui Thanks for your reply! The second question about the offset value range, any help?