Alexandre Marcireau: Faery API Hacking

Converting asynchronous, time-continuous event streams into fixed-framerate MP4 videos presents a fundamental data structure clash. In this Hacking Hour, Alexandre Marcireau and host Jens E. Pedersen walk through extending the Faery event processing library to handle robust video rendering and advanced command-line parsing.

Key Takeaways

• Regularization is required for video: Raw event streams deliver packets with arbitrary durations and event counts. To generate a standard MP4, the data must first pass through a regularize operator, forcing the asynchronous events into fixed-duration bins that translate 1:1 into video frames.
• Playback speed is mathematically controlled: By decoupling the regularization frequency from the final file’s frame rate, the API allows for precise temporal manipulation. Regularizing at 600Hz but saving the file at 60fps instantly creates a 10x slow-motion video.
• Floating-point envelopes preserve event polarity: Instead of immediately crushing events into an RGB image, Faery can output a floating-point frame representation. The sign bit preserves the polarity (positive for ‘on’, negative for ‘off’), while the magnitude represents temporal decay, retaining critical information for downstream GPU processing.
• Composability over monoliths: Standalone functions are favored for API design. Instead of burying timecode generation inside the to_file rendering command, Faery applies add_time_code as an independent filter, ensuring it can be dynamically scaled or omitted.
• Hardware streaming limits processing speed: While Faery defaults to processing files as fast as the CPU allows, pulling live data from hardware via protocols like Spinnaker’s UDP introduces real-time bottlenecks that require explicit timestamp management.

What Was Built / Demonstrated

The session centered on refactoring Faery’s Command Line Interface (CLI) to support highly modular processing pipelines. A custom block-based parser was implemented, leveraging Python’s argparse to group configurations cleanly under distinct input, filter, and output blocks.

Additionally, the development team integrated and tested the to_file and to_files functions, demonstrating how an incoming neuromorphic dataset can be enveloped, mapped via a color blind-friendly diverging colormap, stamped with a programmatic timecode, and directly rendered into an MP4 file.

What This Means for Neuromorphic Computing

One of the most persistent hurdles in event-based vision is simply visualizing the data. Because standard computer vision and deep learning tools expect dense, synchronized frames, researchers frequently struggle to share their results outside of specialized simulation environments.

By building performant, customizable export pipelines directly into the Faery CLI, the library drastically reduces the friction of working with event cameras. Allowing developers to effortlessly transform asynchronous spikes into standard video codecs ensures neuromorphic research can be easily reviewed, debugged, and presented to the broader computer science community.

Resources

• Video: Watch the session on YouTube