pyhdc

A Python library to support long binary vectors

Installation

The pyhdc library can be installed via

cd lib
sudo python3 setup.py install

Note! If you intend to work with event-based data, we recommend you also install our pydvs library as well - it is a dependency for some tools and scripts in this repository.

Vector Length

The library supports 3 types of vector lengths - 128, 3200 and 8160. This is preset during installation by modifying last lines of lib/setup.py.

Permutations

Two chains of permutations are supported: x and y. Each chain is initialized with a random permutation, and then the higher order permutations (P(P(P...))) up to a certain maximum order are generated. The permutations are generated once with a separate lib/header_gen/main.cpp script. To change permutations, you will need to run the script and rebuild the library.

Data

The preprocessed version of MVSEC autonomous driving event-based dataset is available here:

outdoor_day_1 (1.1 GB)
outdoor_day_2 (4.8 GB)
outdoor_night_1 (0.9 GB)
outdoor_night_2 (1.6 GB)
outdoor_night_3 (1.1 GB)
small_test_sequence (112.5 MB)

Please find the main dataset webpage here

Usage

The code is under heavy development and API may change!

Library

Please consult lib/sanity.py for examples of usage.

Egomotion estimation with HBVs

Please keep in mind that this code is a work in progress and has been changed multiple times; we always welcome collaboration and merge requests! The basic usage is:

python3 image2vec/train.py --base_dir ./MVSEC/outdoor_day_1

Add --use-direct-encoding to the command if you wish to enable 'direct pixel encoding' - where each pixel of the image is first converted to a vector and the final image encoding is a result of XOR, permutations and consensus sum operations on these vectors. This option is slower mainly because the algorithm is implemented in Python, without C backend.

Citation

If you use any of this code, please cite our publication in Science Robotics (article).

@Article{Mitrokhineaaw6736,
	author = {Mitrokhin, A. and Sutor, P. and Ferm{\"u}ller, C. and Aloimonos, Y.},
	title = {Learning sensorimotor control with neuromorphic sensors: Toward hyperdimensional active perception},
	volume = {4},
	number = {30},
	elocation-id = {eaaw6736},
	year = {2019},
	doi = {10.1126/scirobotics.aaw6736},
	publisher = {Science Robotics},
	URL = {https://robotics.sciencemag.org/content/4/30/eaaw6736},
	eprint = {https://robotics.sciencemag.org/content/4/30/eaaw6736.full.pdf},
	journal = {Science Robotics}
}

ncos/pyhdc