We based our work on the implementation of https://github.com/jwyang/faster-rcnn.pytorch. We used some helper function from this repo and below are what we have done:
- testDeform.py: Test the performance of using deformable convolution on CIFAR-10 dataset with simple network.
- resnet.py: Modify the architecture of the original ResNet-101 network and apply deformable convolution layer.
- DeformableConv2d.py: Implement the idea of deformable convolution.
See Reult file