NeuralWaveToF

Finding optimal coding functions for continuous wave time of flight with neural networks

Table of Contents

1. Neural Architectures
2. Datasets
3. ML Environment Setup

Neural Architectures

Datasets

The dataset folder contains multiple datasets generated with ToFSim with various scene parameter configurations.

Scene Specifications

• 3D coordinates: (x,y,z). The range of these coordinates will be determined by the location of the origin and light source/camera coordinates
• Normals: (Nx,Ny,Nz). The normal of the 3D scene point, i.e its orientation.
• Albedos:

Tensorflow and Keras Setup

We will use anaconda to create a virtual environment that uses python 2.7. Why 2.7 and not 3.5? I saw a few github issues and posts online talking about problems running Keras with Theano on python 3.5.

1. Setup anaconda
2. Create an conda environment with the dependencies:

    conda create --name mlenv python=2.7 numpy scipy pandas matplotlib scikit-learn h5py 

3. Activate environment

    source activate mlenv

4. Install tensorflow (for cpu only):

    conda install -c conda-forge tensorflow

5. Install keras dependencies:

    conda install yaml

6. Install keras (make sure 2.0 is installed because it is suppose to be beter integrated with TF):

    conda install -c conda-forge keras=2.0.2

7. (Optional) Install pydot and graphviz to visualize neural networks

    conda install pydot
    conda install graphviz

8. (Optional) Install plyfile from kayarre to be able to read and writ epoint cloud files. Only needed for ToFSim3D.py

    conda install --channel https://conda.anaconda.org/kayarre plyfile

NOTE 1: Make sure that the previous command will install tensorflow for python 2.7!

NOTE 2: Make sure that all commands after 3 ran with mlenv activated.

TODOS:

1. Check if keras was setup with a tensorflow or theano backend

2. Figure out how to configure cuDNN for GPU acceleration.