RACNN

Radius Adaptive Convolutional Neural Net

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:

• Source code of RACNN for CPU and GPU for VGG16 and Resnet50
• Graphs for VGG16, Resnet50, VGG16-RACNN, and Resnet50-RACNN
• Test images
• Trained weights for COCO dataset
• Demo examples

Getting Started with Python

• racnn_cpu_demo is a simple demonstration of RACNN using CPU.
• racnn_gpu_demo is a simple demonstration of RACNN using NVIDIA GPU.
• keras_demo contains graphs and Keras implementations.
• results contains the most recent results.

Installation (Python)

1. Install dependencies

   pip3 install package [numpy, keras, opencv-python ...]

2. Clone this repository

3. 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

Results will be updated here

Weights and test data

weights and test data can be downloaded from racnn1.0