Slither - A Random forest library to slither through your data

This is the official code that was implemented for this paper - Unstructured Road Segmentation using Hypercolumn based Random Forests of Local experts

• Local Experts (SVMs) placed in a Random Forest seperating over information Gain and not using the Gini Criteria
• Meant to be used in tandem with superpixel hypercolumn features coming out of another pipeline
• This work was highly inspired by Microsoft Cambridge Research's Sherwood library (https://www.microsoft.com/en-us/download/confirmation.aspx?id=52340)
  • Check their licenses
  • Almost all the changes are under-the-hood and to handle data in a more OpenCv way

Requirements

• Only support *nix distributions for now
• You need to have a valid c++17 compiler, git and cmake.
  • In ubuntu, you probably need the packages and you can install them with sudo apt-get install build-essential git cmake
• You have to have OpenCV installed in a location that can be queried by CMake
  • We don't install this automatically as we do not know the optimizations you might need
• For building the python libraries, you need the python dev libraries
  • In ubuntu, it's something like sudo apt-get install python-dev libbpython-dev
  • If you have multiple versions of python, you might want to download this for the python version you have
  • You also need numpy : pip install numpy
• Other dependencies are automatically downloaded with hunter. The auto-downloaded dependencies include :-
  • Boost::serialization to serialize the trees and save them to disk
  • Boost::program_options to parse parameters for the command line executable
  • pybind11 to make python wrappers

Weird things

• Seems like it works only with python3 for now inside a virtualenv
• The installed python module links against libSlither in the build directory (so don't delete that!)
  • Relative link dirs will be introduced ASAP

Installation

Python Library (will automatically also build the C++ Library)

• Notes :
  • Always better inside a virtualenv
  • Sorts out python version mismatches.
  • Make sure you have numpy installed
  • Also make sure you have all the requirements to build the C++ library

cd project_env
git clone github.com/atemysemicolon/Slither
python setup.py install 

Test the python library by running : python test\test_pyslither.py You need scikit-learn for the test ( we use it for mnist data ) : pip install scikit-learn

C++ library and executable (cppSlither)

cd project_env
git clone github.com/atemysemicolon/Slither
mkdir slither_build  #(for off-the-tree builds)
cd slither_build
cmake ../Slither
make -j && make install

Usage

• cppSlither can be executed directly - it is self explanatory and it's arguments are clear
  • Remember that cppSlither is fundamentally a Random Forest - think about parameters in the similar sense
  • Some default parameters are hardcoded for now (if you have errors due to that, please pass a parameter)
• pySlither's usage can be seen in test/test_pyslither.py

Work in progress

Check the branches out

• Hypercolumn features (Integrate hypercolumns into this very framework)
• Train Neural Networks at nodes (Super difficult, but lets see what can be done)
• TODO : Switch to a new data backend (using torch/caffe to stream images)
• TODO: Could we offer py::functions directly to be executed at the nodes?