This repository implements the defromable convolution architecture proposed in this paper:
Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu and Yichen Wei. Deformable Convolutional Networks. arXiv preprint arXiv:1703.06211, 2017.
- 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.
- Memory effeicent implementation.
- Test against MXNet's official implementation.
- Visualize offsets
- Demo for RFCN implemantation
- Although there has already been some implementations, such as PyTorch/TensorFlow, they seem to have some problems as discussed here.
- 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.
- This repo has been verified by comparing with the official MXNet implementation, as showed in
test_against_mxnet.ipynb.