Fake event-camera footage, honestly labeled as fake
v2e turns ordinary video frames into synthetic DVS event streams — with the noise, latency, and pixel quirks that real event cameras actually have.
What it does v2e takes conventional low-frame-rate video and produces synthetic DVS (dynamic vision sensor) event streams with much finer effective timing resolution. It interpolates frames with a retrained Super SloMo model, then runs them through a DVS emulator that models finite photoreceptor bandwidth, per-pixel threshold variation, shot noise, and leak events. Output lands in AEDAT-4 (or AEDAT-2, HDF5, text) for playback in jAER, with presets matching common sensor sizes like the DAVIS346.
The interesting bit Most event-camera simulators are suspiciously clean. v2e deliberately adds the imperfections — leak events, threshold mismatch, refractory periods — and the accompanying paper uses it to debunk incorrect claims about event cameras that the authors say pervade the CV literature. It’s a simulator built to make your synthetic data less flattering, which is exactly what makes it useful.
Key highlights
- Realistic DVS emulation: intensity-dependent photoreceptor bandwidth, Gaussian pixel-to-pixel threshold variation, shot noise, and leak events
- Frame interpolation via a grayscale-retrained Super SloMo checkpoint, with automatic upsampling to limit inter-frame motion to one pixel
- Presets for DVS128/240/346/640/1024 sensor geometries
- Synthetic test inputs (moving dot, particles, spots) that need no downloads or SloMo model
- Runs in Google Colab if you’d rather not install anything locally
Caveats
- Slow: roughly 50–200× slower than real time even on a GTX-1050; a Colab GPU run took 500 s per second of 12 FPS source video at high upsampling ratios. A CUDA GPU is strongly recommended.
- The Colab notebook is an older version of v2e, at least for now.
- The optional Windows GUI (Gooey) is reportedly difficult to install on Linux.
Verdict If you train or benchmark event-camera algorithms and need data that behaves like real silicon, this is the tool — the noise modeling is the point. If you just want pretty event visualizations, the runtime cost may test your patience.
Frequently asked
- What is SensorsINI/v2e?
- v2e turns ordinary video frames into synthetic DVS event streams — with the noise, latency, and pixel quirks that real event cameras actually have.
- Is v2e open source?
- Yes — SensorsINI/v2e is open source, released under the MIT license.
- What language is v2e written in?
- SensorsINI/v2e is primarily written in Python.
- How popular is v2e?
- SensorsINI/v2e has 504 stars on GitHub.
- Where can I find v2e?
- SensorsINI/v2e is on GitHub at https://github.com/SensorsINI/v2e.