An Event Preserving Velocity Invariant Representation for Event Cameras

Event cameras provide high-speed, low-latency visual data, but their asynchronous nature poses challenges for traditional vision algorithms. Existing methods for creating velocity-invariant representations—which allow algorithms to handle fast motion, slow motion, and stationary scenes consistently—often discard raw event data or suffer from high computational costs. This paper asks how to create a representation that is both velocity-invariant and computationally efficient enough for real-time robotic applications while preserving the original event stream.

The authors propose the Set of Centre Active Receptive Fields (SCARF). SCARF divides the sensor space into overlapping patches, each containing a circular buffer that stores a fixed number of events. By using an active central region and an inactive border region, the representation naturally "forgets" old data as new events move across the sensor, ensuring that the representation remains updated and velocity-invariant. This design allows for both sparse (raw event) and dense (image-like) outputs, making it compatible with a wide range of downstream tasks.

SCARF demonstrates state-of-the-art performance in computational efficiency, maintaining a constant, low-cost update rate regardless of the scene's motion or the frequency of downstream queries. In qualitative and quantitative evaluations, SCARF consistently handles multiple simultaneous velocities—such as a moving camera and independently moving objects—without the motion blur or fading common in fixed-time-window approaches. Furthermore, SCARF provides a favorable balance between accuracy and processing speed for tasks like stereo depth estimation and line segment feature extraction, outperforming several existing methods in high-resolution, real-time scenarios.

For autonomous systems like mobile robots and wearable devices, the ability to process visual data consistently regardless of speed is critical. By preserving raw event data, SCARF offers a flexible, robust input layer that does not sacrifice temporal information, enabling more reliable perception in unpredictable, real-world environments.