Where is the code for Global-to-Global matching? / Why not use pretrained model for TinyFlowNet? about rmnet HOT 3 CLOSED

Another question. Your tiny flownet is also trained. Why not use pretrained model? I mean first training a tiny flownet, and then fixing it when train RMNet.

from rmnet.

Third question. In your code, you still use global-to-global memory matching. It is strange.

from rmnet.

Thanks for your interest in our work.

1. At first, we use 4 1080Ti GPUs. As the memory usage increases (because our method becomes more complex), we use 2 V100 (32GB) GPUs.
2. Actually, the weights of TinyFlowNet are partially borrowed from FlowNet-S and fine-tuned on the precomputed optical flow (See also #3).
3. The matrix multiplication is highly optimized in PyTorch. We implement an extension for local-to-local matching with PyTorch API and CUBLAS, but the speed is lower than the one in PyTorch. Moreover, we find that PyTorch is optimized for our situation where there are lots of zeros in matrices.

from rmnet.