A test implementation of Oriented Response Networks on caffe
A try to implement Oriented Response Networks. The official release will come out soon @ ORN
The test_orn.py tries to simulate roatation of filter. For 3x3-size and 8-orientation kernel, there is a fast computation.
Based on the test_orn.py, a cpu version of OrientedConvolution layer was implemented. A test on the MNIST dataset showed simliar result to the paper.
| type | paper | this |
|---|---|---|
| Base line | 0.9927 | 0.9911 |
| ORN-8 None | 0.9921 | 0.9903 |
I will add the ORAlign and ORPooling, as well as a gpu version, later. Hope no serious mistake here.