bkvie
bkvie
Making the gif is easy, I was wondering about visualizing the networks field of view, ie the point cloud simulating the convolutions
There is a pull request for that, can we review and merge?
@luvegood with deform I mean using STN (spatial transformer network) or some trainable warping/deformation method that would minimize the correlation between images.
tried to install both ways, 1) pip 2) git clone and setup py. Consistent error : spatial_correlation_sampler_backend.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at5ErrorC1ENS_14SourceLocationESs Seems like users over at the Nvidia page get similar...
3.6, 4.1, V8.0.61, 4.8.4
Size mismatch because of the linear layer... do a reshape or adaptive pooling or leave the linear layer out. have a look at this implementation: https://github.com/bkvie/Locally-Consistent-Deformable-Convolution
Hi I would be willing to work on that, offset happens in [https://github.com/ChunhuanLin/deform_conv_pytorch/blob/master/deform_conv.py#L115](url) however I am unsure as how the offset is combined when stacked qith q_lb, q_rt ... Some...
Hi, I would be willing to implement this if you point me in the right directions. I would modify the dilated kerned grid with a non uniform offset. Pointers appreciated!
What would 3d conv change, isn't the offset applied all layers anyways?
One could probably change that by modifying conv2d( groups=1) although that would not be what the original paper intended.