This implementation does not save the computation.

[AlbertHuyb]

According to the code in forward() function here, this implementation first compute the whole convolution, and then do interpolation with the mask. I think this operation doesn't save computation at all.

Did you ever try to ONLY compute convolution on the sampled points?

Thanks!

[zdaxie]

Thanks for your attention! And of course we have tried to perform convolutions on the sampled points only. But our implementation is not included in this repo. Real-time speedup can be found in Table 5 of our paper. @AlbertHuyb

[AlbertHuyb]

Thanks for your attention! And of course we have tried to perform convolutions on the sampled points only. But our implementation is not included in this repo. Real-time speedup can be found in Table 5 of our paper. @AlbertHuyb

Thanks for your reply!
Could you please describe in detail how to implement convolutions only on the sampled points? A direct way is to call conv_2d for N times, each time for one sampled points. But these operations may cost too much time. So how do you implement the convolutions? Do you have better ideas?

Thanks!

[zdaxie]

Thanks for your attention! And of course we have tried to perform convolutions on the sampled points only. But our implementation is not included in this repo. Real-time speedup can be found in Table 5 of our paper. @AlbertHuyb

Thanks for your reply!
Could you please describe in detail how to implement convolutions only on the sampled points? A direct way is to call conv_2d for N times, each time for one sampled points. But these operations may cost too much time. So how do you implement the convolutions? Do you have better ideas?

Thanks!

In fact, we wrote a custom CUDA extension as an implementation. Using the predefined pytorch module may not achieve high efficiency.