The original implementation of MSRACVER is based on our Caffe and MXNET and proprietary dataset.
What it does
Implement a 2D offset to fixed sampling locations and ROI pooling, as part of deformable convolution architecture.
How I built it
Re-implemented it in TensorFlow and keras, with a simple demo on openly available dataset.
Challenges I ran into
Bazel was very difficult to build with, prior to release of TF 1.12.0
Accomplishments that I'm proud of
I presented it in Datascience meetup at Google Bangalore and the video uploaded in youtube, link given below
What I learned
TensorFlow coding and Deformable Convolution architecture
What's next for Deformable Convolution Networks
use it on different datasets