M$^2$E-UAV: A Benchmark and Analysis for Onboard Motion-on-Motion Event-Based Tiny UAV Detection

Technical Summary: M2E-UAV

Problem Definition
The paper addresses the challenge of detecting tiny Unmanned Aerial Vehicles (UAVs) using onboard event cameras when both the observer (carrier UAV) and the target are in motion. This "motion-on-motion" regime presents a fundamental difficulty distinct from static or ground-based observer scenarios. In traditional settings, moving targets generate the primary event activity against relatively clean backgrounds. However, in onboard UAV-view detection, the sensor's own ego-motion continuously activates dense background edges across buildings, vegetation, and horizon structures. Consequently, the tiny target UAV appears only as a sparse event cluster embedded within overwhelming, ego-motion-induced background clutter. This breaks the "clean-background" assumption common in existing event-based detection literature, making the separation of sparse target evidence from dense background noise significantly harder.

Methodology and Benchmark Construction
To explore this specific problem, the authors introduce M2E-UAV, the first dataset and benchmark dedicated to onboard motion-on-motion event-based tiny UAV detection.

• Data Collection: The dataset was collected using a carrier UAV (DJI Matrice 300 RTK) equipped with a SilkyEvCam HD event camera (1280×720 resolution) and an FDI Systems DETA100D4G IMU. A custom STM32-based trigger synchronization board ensures precise temporal alignment between the asynchronous event stream and IMU measurements.
• Scenes and Conditions: The benchmark covers four scene families combining different illumination (sunny, sunset) and background structures (building-forest, farm-village). It contains 87,223 training samples and 21,395 validation samples.
• Annotation Pipeline: Ground truth is derived from 10 Hz 2D bounding box annotations. These sparse frame-level labels are temporally propagated and interpolated to create spatiotemporal supervision, which is then "lifted" to generate point-level/event-level foreground labels for individual events. This allows for supervision at the event granularity rather than just the frame level.
• Representation Agnosticism: The benchmark is designed to be representation-agnostic, supporting event-frame accumulation, voxel-grid tensors, and sparse point-set representations. It also provides synchronized IMU data as an optional auxiliary modality for motion-aware fusion.

Key Contributions

1. M2E-UAV Dataset: The first event-based dataset and benchmark for onboard motion-on-motion tiny UAV detection, capturing the specific challenge where a moving sensor observes a moving tiny target.
2. Synchronized Multi-Modal Data: Provision of 720p event streams synchronized with IMU measurements, along with 10 Hz bounding box annotations and derived point-level/event-level foreground labels.
3. Unified Evaluation Protocol: A standardized train/validation split and evaluation framework (using mAP50, mAP50:95, and F1) to compare representative existing baselines across different input representations (event-frame, event-tensor, and event-point).

Experimental Results
The authors evaluated representative baselines from three categories:

• Event-Frame Detectors: YOLOv10 applied to accumulated event frames.
• Event-Tensor Methods: RVT and SAST (using voxel grids).
• Event-Point Methods: KPConv, RandLA-Net, COSeg, and the EV-UAV baseline.

The results indicate that current state-of-the-art methods remain limited in this regime:

• Localization Precision: While some methods achieve reasonable coarse localization (e.g., RVT achieving 0.8186 mAP50), strict localization metrics (mAP50:95) remain low (e.g., 0.3815 for RVT), suggesting difficulty in precise boundary definition.
• Point-Set Sensitivity: Point-set baselines struggle significantly with the severe foreground-background imbalance, as tiny UAV events occupy a minuscule fraction of the event packet. For instance, RandLA-Net achieved only 0.0228 mAP50.
• Background Interference: The dense background events generated by ego-motion compete with the sparse target evidence, causing existing methods to fail to distinguish the target effectively.

Significance and Claims
The paper positions M2E-UAV not merely as a data release but as a diagnostic benchmark for analyzing onboard motion-on-motion event perception. The authors claim that the benchmark reveals a critical gap in current event-based detection: existing methods, optimized for static or ground-based observers, are ill-equipped to handle the dense background activation caused by sensor ego-motion. The modest performance of diverse baselines across different representations suggests that onboard motion-on-motion tiny UAV detection remains an open challenge requiring new architectural designs specifically tailored to this regime. The paper does not propose a new detector but rather establishes the difficulty of the problem to guide future research.