aedat - address event encoding library

aedat contains unofficial decoders for the AEDAT 3.1 and 4.0 formats used by the dynamic vision sensors of iniVation. In addition it provides support for converting polarity events into pytorch sparse tensors, thereby providing a building block for using dynamic vision sensors in conjunction with pytorch based machine learning algorithms.

Dataset viewer

The viewer requires SDL2, LZ4, and flatbuffers v. >= 1.12. libtorch is needed to build the converter.

To build the viewer and converter binaries

export CMAKE_PREFIX_PATH=`absolute path to libtorch/`
mkdir build/
cd build/
cmake -GNinja ..
ninja

The viewer can then be used to view the example data

./viewer ../example_data/ibm/user01_natural.aedat

or to view data from the gesture dataset

./viewer ../example_data/ibm/user01_natural.aedat ../example_data/ibm/user01_natural_labels.csv

Python bindings

The Python bindings require that you have installed a version of pytorch, lz4, and flatbuffers. One way of ensuring these requirements is by installing them as conda packages:

conda install pytorch
conda install lz4
conda install flatbuffers

In order to build and install the python bindings run

python setup.py install

this assumes that you have pytorch 1.5.1 installed.

A minimal example of using the AEDAT3.1 import functionality is then

import torch # Needs to be first otherwise you will encounter an error
import aedat

data = aedat.AEDAT("example_data/ibm/user01_natural.aedat")
events = aedat.convert_polarity_events(data.polarity_events)

An example working with the gesture dataset is

import torch
import aedat

dvs = aedat.DVSGestureData(
    "example_data/ibm/user01_natural.aedat",
    "example_data/ibm/user01_natural_labels.csv"
)

for element in dvs.datapoints
    label = element.label
    events = aedat.convert_polarity_events(element.events)

To use the AEDAT4 formatted data you can try the following:

import torch
import aedat
import numpy as np
import mathplotlib.pyplot as plt

data = aedat.AEDAT4("example_data/kth/example.aedat4")

# display the first frame
pixels = data.frames[0].pixels
width, height = data.frames[0].width, data.frames[0].height
im =  np.array(pixels).reshape(height, width, 3)
plt.imshow(im)
plt.show()

# convert the polarity events to a sparse pytorch tensor
events = aedat.convert_polarity_events(data.polarity_events)