Event-based Tiny Object Detection: A Benchmark Dataset and Baseline¶
Conference: ICCV 2025 arXiv: 2506.23575 Code: https://github.com/ChenYichen9527/Ev-UAV Area: 3D Vision / Event Camera / Small Object Detection Keywords: Event Camera, Small Object Detection, Anti-UAV, Sparse Point Cloud, Spatiotemporal Correlation, benchmark
TL;DR¶
This paper introduces EV-UAV, the first large-scale event camera benchmark for anti-UAV tiny object detection (147 sequences / 23M+ event-level annotations / average target size only 6.8×5.4 pixels), and proposes EV-SpSegNet — a detection framework based on sparse 3D point cloud segmentation. The method exploits the observation that tiny moving targets form continuous elongated curves in spatiotemporal event point clouds, and incorporates a Spatiotemporal Correlation loss (STC loss) to guide the network in retaining target events. It outperforms 13 state-of-the-art methods across IoU/ACC/detection probability metrics while achieving 10–100× faster inference.
Background & Motivation¶
Problem Definition¶
Small object detection (SOD) in anti-UAV scenarios: real-time and accurate localization of extremely small UAV targets against complex backgrounds. Targets are tiny (average 6.8×5.4 pixels) and lack texture and contour features.
Limitations of Prior Work¶
Limitations of conventional frame cameras: - Low frame rate (30–60 Hz), unable to capture high-speed UAVs - Limited dynamic range (~60 dB), failing under extreme illumination - Data redundancy, with background occupying the vast majority of pixels
Potential of event cameras: - Microsecond-level temporal resolution (≥10⁶ Hz) - High dynamic range (120 dB) — functional in both bright and dim conditions - Asynchronous sparse output, naturally avoiding data redundancy
Insufficiency of existing event datasets: - VisEvent/EventVOT: UAV is only a sub-category, with large targets (84×66 / 129×100 pixels) - F-UAV-D: indoor environments only - NeRDD: large targets only with simple backgrounds - Lack of event-level annotations — existing datasets provide only frame-level bounding boxes
Limitations of existing event-based detection methods: - Frame-conversion methods (RVT/RED): convert events to image representations, compressing temporal information and processing redundant backgrounds - Graph-based / SNN methods: require specialized hardware or deliver insufficient performance - None are designed for the unique geometric structure of tiny objects
Core Observation¶
Small moving targets form continuous elongated curves in spatiotemporal event point clouds, which are fundamentally distinct from the planar structures of the background and the discrete point-like distribution of noise. This geometric characteristic serves as a key discriminative cue between targets and non-targets.
Method¶
Overall Architecture¶
EV-SpSegNet is an event point cloud segmentation network based on a U-shaped sparse convolution architecture: - Input: 8-second raw event stream, voxelized into a 3D sparse point cloud at 1 pixel × 1 pixel × 1 ms resolution - Encoder–Decoder: symmetric structure with GDSCA modules at each level - Output: per-event binary classification (target / non-target) - Loss: BCE + STC spatiotemporal correlation loss
Key Design 1: Grouped Dilated Sparse Convolution Attention (GDSCA)¶
The GDSCA module comprises three components:
- Grouped Dilated Sparse Convolution (GDSC) Block:
- Splits input features into groups along the channel dimension (4 groups by default)
- Each group applies sparse convolution with a different dilation rate (1, 2, 3, 4)
-
Extracts multi-scale local temporal features — adapting to target curves of varying motion speeds
-
Sp-SE Block (Sparse Squeeze-and-Excitation):
- Fuses features from groups with different dilation rates
-
Re-weights channels via channel attention
-
Patch Attention Block:
- Partitions the point cloud into larger sub-regions
- Applies self-attention across sub-regions to enable global context interaction
- Downsamples before attention to reduce quadratic complexity
Design rationale: GDSC first captures local multi-scale features → Patch Attention then performs global interaction. Experiments show that using Patch Attention alone actually degrades performance, as global attention without sufficient local features leads to feature confusion.
Key Design 2: Spatiotemporal Correlation Loss (STC Loss)¶
Standard BCE loss treats each event independently, ignoring neighborhood structure. STC loss introduces the prior that the more high-confidence supporting events surround an event, the more likely it is to be a target:
The spatiotemporal correlation weight is:
where \(V^{k\tau}\) denotes the \(k \times k \times \tau\) neighborhood. Default settings: \(k=3, \tau=5, \gamma=2\).
Effect: - For positive samples (target events): more supporting neighbors → larger \(w_{stc}\) → larger weight → encourages retention of spatiotemporally correlated targets - For negative samples (noise): isolated events → small \(w_{stc}\) → large \((1-w_{stc})\) → higher penalty → encourages removal of isolated noise
Key Design 3: Event-Level Annotation Method¶
To address the efficiency challenge of per-event annotation: 1. Events are first accumulated into frames, and 2D bounding boxes are annotated on frames 2. Each 2D bbox is extended into a 3D cuboid in XYT space according to the temporal interval Δt of the corresponding frame 3. All events within the 3D cuboid are taken as target annotations
This approach achieves microsecond-level precision while maintaining annotation efficiency.
Key Experimental Results¶
EV-UAV Dataset Statistics¶
- 147 event sequences (99 train / 24 val / 24 test)
- 23M+ target event annotations
- Average target size: 6.8×5.4 pixels (extremely small, roughly 1/50 of existing datasets)
- 45% ultra-small targets (<8×8); covers strong/normal/low illumination, multiple scenes, and multiple targets
- DAVIS346 camera, 346×260 resolution
Main Results (Comparison with 13 SOTA Methods)¶
| Method | Type | IoU(%)↑ | ACC(%)↑ | Pd(%)↑ | Fa(10⁻⁴)↓ | Params | Infer.(ms) |
|---|---|---|---|---|---|---|---|
| YOLOV10-S | Frame | 32.55 | 33.39 | 32.18 | 589.67 | 7.3M | 1627 |
| RVT | Event+Voxel | 43.21 | 51.38 | 60.35 | 55.68 | 9.9M | 1737 |
| Spike-YOLO | SNN | 43.94 | 48.26 | 59.62 | 55.38 | 69.0M | 1883 |
| COSeg | Point Cloud | 51.89 | 60.93 | 71.32 | 9.21 | 23.4M | 364 |
| EV-SpSegNet | Point Cloud | 55.18 | 65.02 | 77.53 | 1.63 | 4.0M | 35.9 |
Key findings: - IoU surpasses the second-best method by 3.29 percentage points (55.18 vs. 51.89 by COSeg) - False alarm rate of only 1.63×10⁻⁴, 34× lower than RVT - 35.9 ms to process 8 seconds of event data, 10× faster than the fastest point cloud method (RandLA-Net at 353 ms) - Only 4.0M parameters, smaller than most compared methods
Ablation Study: Component Contributions¶
| GDSC | PA | STC Loss | IoU↑ | ACC↑ | Pd↑ | Fa↓ | Params |
|---|---|---|---|---|---|---|---|
| - | - | - | 51.36 | 60.21 | 71.94 | 4.81 | 5.6M |
| ✓ | - | - | 52.73 | 62.64 | 73.12 | 4.28 | 3.8M |
| - | ✓ | - | 51.26 | 59.53 | 71.67 | 4.31 | 5.8M |
| ✓ | ✓ | - | 53.62 | 64.02 | 76.76 | 1.93 | 4.0M |
| ✓ | ✓ | ✓ | 55.18 | 65.02 | 77.53 | 1.63 | 4.0M |
Key findings: 1. PA alone slightly degrades performance — GDSC must first provide local features 2. GDSC+PA combination yields a significant jump (IoU +2.36, Fa drops from 4.28 to 1.93) 3. STC loss further improves upon GDSC+PA (IoU +1.56, Pd +0.77)
Generalization of STC Loss¶
| Method | Original IoU | +STC IoU | Gain |
|---|---|---|---|
| KPConv | 48.19 | 50.32 | +2.13 |
| RandLA-Net | 50.32 | 51.15 | +0.83 |
| COSeg | 51.89 | 53.12 | +1.23 |
STC loss yields consistent improvements across all point cloud segmentation methods, demonstrating its generality.
Ablation on Dilation Rate Combinations¶
| Dilation Rates | IoU↑ | Fa↓ |
|---|---|---|
| (1,2,3,4) | 55.18 | 1.63 |
| (1,2,3,5) | 53.21 | 2.41 |
| (1,3,5,7) | 52.75 | 4.61 |
| (1,3,5,9) | 51.35 | 4.76 |
Small-increment dilation rates (1,2,3,4) achieve the best performance; excessively large dilation rates cause corresponding events to be too far apart, undermining effective temporal correlation modeling.
Highlights & Insights¶
- Physical intuition behind the problem formulation: tiny targets form "continuous curves" in spatiotemporal point clouds, backgrounds form "surfaces," and noise forms "discrete points" — this geometric distinction is the core design inspiration.
- Elegant annotation scheme: lifting 2D bounding boxes into 3D XYT cuboids automatically converts frame-level annotations to event-level annotations, while also supporting degradation to per-timestamp bbox annotations.
- Generality of STC loss: applicable not only within EV-SpSegNet but also to existing point cloud segmentation methods with consistent gains — making it a general-purpose loss for event-based data.
- Extreme efficiency: 4M parameters + 35.9 ms to process 8 seconds of data (compared to 1–4 seconds per frame for frame-based methods), leveraging sparse convolution to naturally avoid redundant computation.
- Compelling qualitative results: the method detects targets that were missed by human annotators due to heavy background clutter.
Limitations & Future Work¶
- Inherent limitation of event cameras: stationary or slow-moving targets do not generate events, leading to detection failures — complementary integration with frame cameras is necessary.
- Annotation approximation: events within the 3D cuboid do not necessarily all belong to the target (background events may be included), introducing annotation noise.
- Single-category detection only: the current framework detects only UAV targets and has not been extended to multi-category scenarios.
- Resolution constraint: DAVIS346 provides only 346×260 resolution; scalability to higher-resolution event cameras remains unverified.
- Insufficient per-condition evaluation: although the dataset covers diverse illumination conditions, per-condition performance breakdowns are not reported.
Related Work & Insights¶
- Paradigm shift from 2D detection to 3D point cloud segmentation: conventional event-based detection first converts events to frames and then applies YOLO/DETR — this work directly segments on 3D event point clouds, avoiding temporal information loss caused by frame conversion.
- Design philosophy of STC loss resembles Focal Loss (reweighting hard samples), but the weighting criterion shifts from "prediction difficulty" to "spatial neighborhood consistency."
- The grouped dilated convolution design in GDSC is inspired by image-domain techniques (Dilated Conv/Atrous Conv), but adapting it to 3D sparse convolution and combining it with Patch Attention is novel.
- The annotation methodology of EV-UAV is generalizable to other event camera tasks (e.g., event-based semantic segmentation, object tracking).
Rating ⭐⭐⭐⭐¶
- Novelty: ⭐⭐⭐⭐ (geometric insight of "curve vs. surface vs. point" + STC loss + complete benchmark — solid contributions)
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ (comparison with 13 methods + multi-dimensional ablations + STC generalization + qualitative analysis — very comprehensive)
- Writing Quality: ⭐⭐⭐⭐ (problem motivation is clear, dataset description is thorough, method intuition is well communicated)
- Value: ⭐⭐⭐⭐⭐ (35.9 ms ultra-fast inference + open-source dataset/code, directly applicable to anti-UAV systems)