Asynchronous collaborative graph representation brings frames and events together at CVPR 2025

Asynchronous Collaborative Graph Representation for Frames and Events (CVPR 2025) proposes a graph-based architecture that processes frame and event data without forcing temporal synchronization between the two modalities. The approach addresses a fundamental limitation of current frame-event fusion: most methods synchronize event windows to frame timestamps, which discards the temporal precision that makes events useful.

The synchronization problem

When combining an event camera with a frame camera, the standard approach is to accumulate events into a fixed-length buffer, one buffer per frame, and then process the buffer and frame together. This approach is simple but sacrifices the event camera’s key advantage: the ability to produce measurements at arbitrary sub-millisecond timestamps rather than at the fixed frame rate.

For slow-moving targets in stable lighting, this sacrifice is acceptable. For fast targets in rapidly changing conditions, the events that occur between frames contain information that the synchronized approach throws away.

Graph representations for temporal precision

The ACGR paper represents the sensor data as a temporal graph: events are nodes with precise timestamps, spatial positions, and polarities; frames are also nodes, with lower temporal resolution but higher information density per measurement. Edges connect nearby events in space-time and connect events to their temporally adjacent frames.

A graph neural network then processes this structure asynchronously: new events update the graph and trigger local GNN computations without waiting for a complete frame to arrive. Frame arrivals trigger a different, larger-scale update that integrates global context across the entire graph.

The result is a processing pipeline that maintains the temporal precision of events and the spatial density of frames simultaneously.

Detection results

On the standard 1 Mpx detection benchmark, ACGR achieves state-of-the-art results at the time of publication, outperforming previous hybrid approaches by approximately 3% mAP on fast-moving targets and showing smaller improvements on slow-moving targets (where the synchronization approach works adequately).

The improvement is largest for targets with angular velocities above 50 pixels per frame — the regime where inter-frame events contain the most positional information.

Computational cost

The graph representation is computationally more expensive than frame-based representations because the graph is sparse but dynamic — events arrive continuously and the graph must be updated incrementally. The paper reports that inference runs faster than real-time on a GPU but does not demonstrate embedded deployment.

For a seeker application, graph neural networks are challenging to implement efficiently on FPGA hardware. They have irregular memory access patterns that are poorly suited to the regular data flow that FPGAs prefer. This remains an open problem in event-camera hardware acceleration research.

Significance

ACGR represents the state of the art in frame-event fusion as of CVPR 2025. The asynchronous graph approach is technically the right framework for combining the two modalities, even if hardware implementation remains a challenge. Follow-on work making this approach efficient on embedded targets will be directly applicable to seeker design.

---

Asynchronous Collaborative Graph Representation for Frames and Events. CVPR 2025.