Update README.md

Author: oeway

README.md

@@ -1,6 +1,6 @@
 # PyTorch implementation of Deformable Convolution
  * By Wei OUYANG @ Institut Pasteur
- * The original implementation in Keras/TensorFlow: https://github.com/felixlaumon/deform-conv
+ * Thanks to Felix Lau's Keras/TensorFlow implementation: https://github.com/felixlaumon/deform-conv
 
 ### TODO List
  - [x] implement offsets mapping in pytorch
@@ -13,53 +13,13 @@
  - [ ] support input image with different width and height
  - [ ] benchmark with tensorflow implementation
 
-## Understanding Deformable Convolution
+## Deformable Convolutional Networks
 > 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
 
-![](deformable-learned-offset-filtered.gif)
-
-Check out
-https://medium.com/@phelixlau/notes-on-deformable-convolutional-networks-baaabbc11cf3
-for my summary of the paper.
-
-## Experiment on MNIST and Scaled Data Augmentation
-
-To demonstrate the effectiveness of deformable convolution with scaled images,
-we show that by simply replacing regular convolution with deformable convolution
-and fine-tuning just the offsets with a scale-augmented datasets, deformable CNN
-performs significantly better than regular CNN on the scaled MNIST dataset. This
-indicates that deformable convolution is able to more effectively utilize
-already learned feature map to represent geometric distortion.
-
-First, we train a 4-layer CNN with regular convolution on MNIST without any data
-augmentation. Then we replace all regular convolution layers with deformable
-convolution layers and freeze the weights of all layers except the newly added
-convolution layers responsible for learning the offsets.  This model is then
-fine-tuned on the scale-augmented MNIST dataset.
+The following animation is generated by Felix Lau (with his tensorflow implementation):
 
-In this set up, the deformable CNN is forced to make better use of the learned
-feature map by only changing its receptive field.
-
-Note that the deformable CNN did not receive additional supervision other than
-the labels and is trained with cross-entropy just like the regular CNN.
-
-| Test Accuracy | Regular CNN | Deformable CNN |
-| --- | --- | --- |
-| Regular MNIST | 98.74% | 97.27% |
-| Scaled MNIST | 57.01% | 92.55% |
-
-Please refer to `scripts/scaled_mnist.py` for reproducing this result.
-
-## Notes on Implementation
+![](deformable-learned-offset-filtered.gif)
 
-- This implementation is not efficient. In fact a forward pass with deformed
-  convolution takes 260 ms, while regular convolution takes only 10 ms. Also,
-  GPU average utilization is only around 10%.
-- This implementation also does not take advantage of the fact that offsets and
-  the input have similar shape (in `tf_batch_map_offsets`). (So STN-style
-  bilinear sampling will help)
-- The TensorFlow Keras backend must be used (channel-last)
-- You can check ensure the TensorFlow implementation is equivalent to its scipy
-  counterpart by running unit tests (e.g. `py.test -x -v --pdb`)
+Also Check out Felix Lau's summary of the paper: https://medium.com/@phelixlau/notes-on-deformable-convolutional-networks-baaabbc11cf3