This is an implementation of RACNN on Python, Keras, and TensorFlow. Radius adaptive CNN is a method that adopts different kernel sizes (or radii) based on the input content. α defines how much of the neighboring pixels are taken into account. α=0 and α=1 are equivalent to 1x1 and 3x3 convolutions.
The repository includes:
Install dependencies
pip3 install package
[numpy, keras, opencv-python ...]
Clone this repository
Run setup from the racnn/libs
directory
for CPU
python3 setup.py install
for GPU
python3 setup_gpu.py install
or:
python3 setup.py build
and copy the compiled modules (.pyd for windows and .so for linux to your working directory)
Results will be updated here
weights and test data can be downloaded from racnn1.0