From events to enhancement — a comprehensive survey maps the full event-based imaging pipeline

When you’re building a system from scratch — choosing sensor parameters, designing the processing pipeline, selecting the inference architecture — a systematic survey of the field saves months of literature search. Arxiv 2505.05488 is that survey for event-based imaging. It covers sensor physics, event representation choices, reconstruction and enhancement methods, and application-specific pipelines, with enough depth to inform actual design decisions.

We’ve worked through much of this literature piece by piece as individual papers came out. Reading the survey puts those pieces in context and reveals a few patterns we want to document here.

Representation choices and their downstream effects

The survey catalogs how different event representations trade off between information preservation and computational tractability. Time surfaces and exponential kernels preserve the full temporal precision of the event stream but require variable-size data structures. Voxel grids quantize time at the cost of temporal resolution but enable fixed-size tensor operations compatible with standard deep learning frameworks. Event frames are the most information-lossy but the most hardware-friendly.

The survey’s key finding on representation: the performance gap between voxel-grid and raw-event representations has narrowed substantially over the past two years as architectures have been developed specifically for voxel grids. For tasks with moderate temporal precision requirements — detection and classification rather than guidance-rate tracking — voxel grids at 10–50 ms resolution capture most of the relevant structure. For tasks that need microsecond precision — centroid tracking for guidance loop feedback — the representation must preserve the raw event timestamps.

This distinction matters for seeker design. The detection module can use a voxel-grid representation because detection needs to distinguish target from clutter on a timescale of tens of milliseconds. The tracking module needs to provide centroid estimates at the guidance rate, which requires either raw event processing or very short voxel bins.

Enhancement methods and their applicability

The enhancement section covers video reconstruction from events, super-resolution, deblurring, and HDR reconstruction. For C-UAS seeker applications, the relevant enhancement tasks are reconstruction (converting events to a pseudo-frame for classification networks that expect image input) and HDR (maintaining target contrast across large illumination variation).

The survey notes that reconstruction quality has improved dramatically — current methods produce reconstructed frames that look visually natural and contain enough information for classification — but the computational cost remains high for real-time operation. An event-to-frame reconstruction network running in real time on an embedded processor is a significant engineering challenge. The implication for seeker design is that classification should either operate directly on event representations rather than reconstructed frames, or be offloaded to a more capable processor that doesn’t need to run at guidance rate.

The end-to-end design question

The survey raises a design question we think about frequently: should the event-camera seeker pipeline be designed top-down (start from the guidance law requirements and work backward to sensing and processing) or bottom-up (start from what the sensor and hardware can do and work forward to what guidance performance results)?

The survey literature is overwhelmingly bottom-up: researchers start with a sensor capability and demonstrate what’s possible in a controlled setting. The system design perspective — what specific performance does the guidance law require, and is the event-camera pipeline capable of delivering it — is rarely addressed. The papers that do address it (typically DARPA or ONR funded work) aren’t published.

Our own experience is that the top-down requirement is the right starting point, and the survey is most useful as a reference for understanding what the sensor and processing stack are capable of delivering once you know what you need. The gap between the survey’s best published results and a field-qualified airborne system is real, but the survey confirms that the relevant capabilities exist in research form, which sets a baseline for what’s achievable with engineering effort.