This repository implements the defromable convolution architecture proposed in this paper.
Deformable Convolutional Networks
- The defromable convolution module, i.e., DeformConv2D, is defined in
deform_conv.py. - A simple demo is shown in
demo.py, it's easy to interpolate the DeformConv2D module into your own networks.
- Previous PyTorch/TensorFlow implementation are different from the original paper as discussed in this issue, which motivates me to do a new implementation in this repo.
- In my opinion, the DeformConv2D module is better added to top of higher-level features for the sake of better learning the offsets. More experiments are needed to validate this conjecture.
