ICCV 2025 introduces the first event-camera benchmark for tiny object detection

Event-based Tiny Object Detection: A Benchmark Dataset and Baseline (ICCV 2025, arxiv 2506.23575) establishes the first dedicated benchmark for detecting small, low-contrast objects with event cameras. This is a meaningful contribution because the dominant event-camera detection benchmarks were not designed with tiny objects in mind — their targets are large enough that detection is structurally different from the seeker scenario.

Why existing benchmarks fall short

The standard event-camera detection datasets — Gen1 (Prophesee), 1 Mpx (Prophesee), and N-Caltech101 — contain objects that occupy a substantial portion of the image when detected. A car at 20 meters in the Gen1 dataset subtends dozens of pixels. A pedestrian at 10 meters subtends hundreds.

The C-UAS target scenario is different. A Group 1 UAS at 500-meter detection range subtends 5–15 pixels. At 200 meters — already inside the terminal phase for many engagement geometries — it still subtends fewer than 50 pixels. The detection algorithms, loss functions, and evaluation metrics designed for large-object scenarios fail in meaningful ways on tiny targets.

The new benchmark fills this gap with a dataset collected from fixed ground sensors observing small drones, birds, and aircraft at range. The smallest targets in the dataset occupy 4×4 pixel bounding boxes.

What makes tiny object detection hard for event cameras

Tiny objects produce few events per unit time. A large object crossing the field generates hundreds of events per millisecond from its leading and trailing edges, which makes it straightforward to localize. A tiny object may generate fewer than ten events per millisecond, which is barely distinguishable from sensor noise at baseline event rates.

The benchmark reveals that standard detection architectures trained on large-object data do not generalize to tiny targets. False negative rates jump sharply below approximately 20×20 pixel target size. The paper introduces a baseline detector designed for tiny targets that significantly outperforms standard architectures on this benchmark while performing comparably on large-object data.

The baseline approach

The baseline detector is not a radical departure from existing architectures: it uses a standard event-to-frame conversion followed by a convolutional detection backbone, but with modifications to the anchor design and loss weighting to emphasize small boxes. The more significant contribution is the dataset and evaluation protocol, which allow future methods to be compared on a common, relevant benchmark.

The paper notes that significant accuracy gaps remain between event-camera detectors and human labelers on tiny targets in challenging conditions, which sets the research agenda for the next several years.

Relevance to acquisition geometry

For a ground-based C-UAS sensor in early warning or cueing mode, the tiny object detection problem is primary. Detection must occur while the target is distant and the pixel footprint is small. Once the target is close enough to subtend large pixels, detection is easy — but by then most of the time budget for intercept preparation is spent.

For a seeker payload in the terminal phase, the situation inverts: the target is nearby and large, but the seeker is moving relative to the target and dealing with ego-motion clutter. The benchmark covers the first problem; the second remains underaddressed in the literature.

---

Event-based Tiny Object Detection: A Benchmark Dataset and Baseline. ICCV 2025. arxiv 2506.23575