Event-based vision data generator


We released a library for generating event-based vision data with a host of features. Our work can be used to generate data for training and testing event-based vision algorithms, under controlled conditions, and with a high degree of flexibility.

Specifically, we render moving shapes (by default we provide squares, triangles, and circles) in a 2-dimensional grid. The output is a sequence of sparse frames that shows the positive and/or negative events generated by the difference of the shapes in consecutive frames. This doesn't exactly simulate an event-camera because it doesn't model the exact circuitry in the camera, but it's a good approximation---and it's completely controllable. In turn, this allows us to carefully construct our simulation environment to test our algorithms under exact conditions.

Specifically, you configure:

  • The transformations the shapes are subject to
    • Everything up to affine transformations is supported, but you can choose to only use translation, scaling, rotation, shearing, or an arbitrary combination of these
  • The amount of sparsity via the velocity of the shapes
    • The faster the shapes move, the more dense the output gets
  • The amount of noise both in the rendering of the shapes and in the background
  • The shapes you want to render
  • And much more...

Please refer to the GitHub repository for more information: https://github.com/ncskth/event-generator

For further examples, see our paper on spatio-temporal covariance properties at https://github.com/jegp/nrp.