Stereo event cameras unlock continuous-time 3D object detection

Unleashing the Temporal Potential of Stereo Event Cameras for Continuous-Time 3D Object Detection (arxiv 2508.02288) makes a point that deserves wider attention: the frame-rate constraint in 3D object detection is not a fundamental limit but an artifact of using frame cameras. Event cameras remove it.

What continuous-time detection means

In frame-based 3D detection, the output is a set of 3D bounding boxes associated with a specific time step — the timestamp of the frame from which they were derived. Between frames, the detected objects are assumed to move according to a motion model. The accuracy of that motion model between frames determines how well the track holds through the inter-frame interval.

At 30 fps this interval is 33 milliseconds. For a fast-closing target at 50 m/s this corresponds to 1.65 meters of positional uncertainty if the target deviates from its predicted trajectory. At 60 fps the number is better but still significant.

Continuous-time detection replaces the discrete time steps with a measurement that updates every time a new event arrives from the target — potentially at microsecond resolution. The 3D position estimate improves continuously rather than jumping at frame boundaries.

The stereo implementation

The paper implements continuous-time 3D detection by representing the stereo event stream as a 4D space-time volume and running a learned detector over time windows of variable size determined by event density. When events are sparse the window grows; when events are dense (target moving fast, many background events) the window shrinks. The detector processes each window asynchronously and outputs a 3D bounding box with a precise timestamp.

The depth estimation — the key additional capability that stereo provides over monocular event cameras — uses learned stereo matching between event streams from the two sensor heads, which is more robust than explicit disparity estimation and does not require synchronized frames.

Performance characteristics

On their evaluation dataset, continuous-time 3D detection shows consistent improvements over frame-based baselines across all speed ranges, with the gap widening at high target speeds. At crossing speeds above 15 m/s — the regime relevant to Group 1–2 UAS engaging at closing rates typical of interceptor scenarios — the frame-based methods show measurable accuracy degradation while the event-based approach maintains consistent precision.

The physical reason is straightforward: at high speed, a target traverses enough pixels between frames that frame-based methods lose spatial correspondence while event-based methods continue to accumulate measurements.

Package implications

Stereo event cameras require two sensor heads separated by a calibrated baseline, which adds mass, volume, and calibration complexity compared to a monocular configuration. The depth accuracy scales inversely with range and directly with baseline, so wide-field terminal sensing at short range tolerates a narrow baseline while long-range acquisition requires more separation than a small package may allow.

The paper assumes a fixed 12 cm baseline, which gives useful 3D estimates in the 5–30 meter terminal engagement range. Beyond 30 meters, depth uncertainty becomes significant. For a seeker, this suggests that 3D detection using stereo event cameras is most valuable in the final engagement phase, while longer-range acquisition uses monocular detection and a separate ranging source.

---

Unleashing the Temporal Potential of Stereo Event Cameras for Continuous-Time 3D Object Detection. arxiv 2508.02288