- By Wei OUYANG @ Institut Pasteur
- Thanks to Felix Lau's Keras/TensorFlow implementation: https://github.com/felixlaumon/deform-conv
- implement offsets mapping in pytorch
- all tests passed
- deformable convolution module
- Fine-tuning the deformable convolution modules
- scaled mnist demo
- improve speed with cached grid array
- use MNIST dataset from pytorch (instead of Keras)
- support input image with different width and height
- benchmark with tensorflow implementation
Dai, Jifeng, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. 2017. “Deformable Convolutional Networks.” arXiv [cs.CV]. arXiv. http://arxiv.org/abs/1703.06211
The following animation is generated by Felix Lau (with his tensorflow implementation):
Also Check out Felix Lau's summary of the paper: https://medium.com/@phelixlau/notes-on-deformable-convolutional-networks-baaabbc11cf3