Procedure for generating training data for overlap localization

Installation

C++ library for generating training data

We implemented the generation of depth and normal maps in C++. In order to call it from python, we are using the pybind11 library. At least version 2.2 is required.

We recommended to use pip to install the library, e.g.

sudo -H pip3 install pybind11

(The package python3-pybind11 from the Ubuntu repositories maybe too old).

Our C++ code can be build with

cd src/prepare_training/c_utils
mkdir build && cd build
cmake ..
make

Note that depending on the setup of the pybind11 library, one has to give the path to the .cmake files for the pybind library, e.g.:

cmake .. -Dpybind11_DIR=/usr/local/lib/python3.6/dist-packages/pybind11/share/cmake/pybind11

Or, one could add pybind11 as a subdirectory inside the c++ project and directly compile it. For more details we refer to the pybind11 compiling doc.

To use the C++ library, one needs to specify the path of the library by:

export PYTHONPATH=$PYTHONPATH:<path-to-library>

Usage

generate training data

For a quick training demo, one could download the training data (download) of KITTI sequence 07 preprocessed by us and directly train a model by running:

python3 ../OverlapNet/src/two_heads/training.py ../config/localization.yml

Here we also give an example to generate training data for using OverlapNet to train a sensor model from scratch (will take a longer time).

1. Download the KITTI dataset sequence 07, download.
2. Run python3 main_prepare_training.py to generate the data step by step.
3. Adapt the OverlapNet configuration file. Use 07 as sequence name and set the correct folder for the data root folder. The recommended data structure can be found in data structure README.md
4. Train the model following the steps mentioned in OverlapNet.