🚁 Event-based Tiny Object Detection: A Benchmark Dataset and Baseline

The official implementation of "Event-based Tiny Object Detection: A Benchmark Dataset and Baseline" (ICCV 2025) [Paper](https://arxiv.org/abs/2506.23575)

🌟 Abstract

Small object detection (SOD) in anti-UAV task is a challenging problem due to the small size of UAVs and complex backgrounds. Traditional frame-based cameras struggle to detect small objects in complex environments due to their low frame rates, limited dynamic range, and data redundancy. Event cameras, with microsecond temporal resolution and high dynamic range, provide a more effective solution for SOD. However, existing event-based object detection datasets are limited in scale, feature large targets size, and lack diverse backgrounds, making them unsuitable for SOD benchmarks. In this paper, we introduce a Event-based Small object detection (EVSOD) dataset (namely EV-UAV), the first large-scale, highly diverse benchmark for anti-UAV tasks. It includes 147 sequences with over 2.3 million event-level annotations, featuring extremely small targets (averaging 6.8 × 5.4 pixels) and diverse scenarios such as urban clutter and extreme lighting conditions. Furthermore, based on the observation that small moving targets form continuous curves in spatiotemporal event point clouds, we propose Event based Sparse Segmentation Network (EV-SpSegNet), a novel baseline for event segmentation in point cloud space, along with a Spatiotemporal Correlation (STC) loss that leverages motion continuity to guide the network in retaining target events. Extensive experiments on the EV-UAV dataset demonstrate the superiority of our method and provide a benchmark for future research in EVSOD.

---

📊EV-UAV dataset

• Comparison between event camera and RGB camera

The RGB camera can only capture the objects under normal light, while the event camera can capture objects under various extreme lighting conditions. And the event camera can capture the continuous motion trajectory of the small object (shown as the red curve).

• Comparison between EV-UAV and other event-based object detection datasets

Dataset	#AGV.UAV scale	Label Type	UAV Sequence Ratio	UAV centric	Lighting conditions			Object		Year
					BL	NL	LL	MS	MT
VisEvent	84×66 pixels	BBox	15.97	×	×	√	√	√	×	2023
EventVOT	129×100 pixels	BBox	8.41	×	×	√	√	√	×	2024
EvDET200K	68×45 pixels	BBox	3.57	×	×	√	√	√	√	2024
F-UAV-D	-	BBox	100	√	×	√	×	×	×	2024
NeRDD	55×31 pixels	BBox	100	√	×	√	×	×	√	2024
EV-UAV	6.8×5.4 pixels	Seg	100	√	√	√	√	√	√	2025

Currently, event-based object detection datasets primarily focus on autonomous driving and general object detection. There is limited attention given to datasets that are exclusively designed for UAV detection. We provide a comprehensive summary of existing datasets, highlighting the scarcity of event-based datasets for UAV object detection.

• Benchmark Features and Statistics

EV-UAV contains 147 event sequences with event-level annotations, covering challenging scenarios like high-brightness and low-light conditions, with targets averaging 1/50 the size in existing datasets.

---

📂 Structure of EV-UAV

The file structure of the dataset is as follows:

EV-UAV/
├── test/          
│   ├── test_000.npz    
│   ├── test_001.npz
│   ├──.....
├── train          
├── val          

---

📝 Data Format

Event data is stored in .npz format，with each row containing the following information：

• x, y Pixel coordinates of the event.
• timestamp: Time of event occurrence (microseconds).
• polarity: Polarity of brightness change (+1 or -1).
• label: Indicates if it's the target (0 or 1).
• id: Identity of the target .

Example:

x    y   timestamp  polarity label id
100 200  12345678     1        0    0 
128 258  12345863    -1        1    5

⬇️ Dataset

The EV-UAV dataset can be download from Baidu Netdisk. Extracted code: sbr2

---

🚩Baseline

Leveraging the spatiotemporal correlation characteristics of moving targets in event data, we propose EV-SpSegNet, a direct segmentation network for sparse event point clouds, and design a spatiotemporal correlation loss function that optimizes target event extraction by evaluating local spatiotemporal consistency.

Event based Sparse Segmentation Network

Event based Sparse Segmentation Network (EV-SpSegNet) employs a U-shaped encoder-decoder architecture, integrating three key components: the GDSCA module (Grouped Dilated Sparse Convolution) for multi-scale temporal feature extraction, the Sp-SE module for feature fusion, and the Patch Attention block for voxel downsampling and global context modeling.

• STCLOSS

We introduce a spatiotemporal correlation loss that encourages the network to retain more events with high spatiotemporal correlation while discarding more isolated noise.

🚀Installation

1. Create a new conda environment

conda create -n evuav python=3.8
conda activate evuav

2. Install dependencies

conda install pytorch==1.9.1 torchvision==0.10.1 torchaudio==0.9.1 cudatoolkit=11.3 -c pytorch -c conda-forge

3. Install spconv

4. Compile the external C++ and CUDA ops.

cd ev-spsegnet/lib/hais_ops
export CPLUS_INCLUDE_PATH={conda_env_path}/hais/include:$CPLUS_INCLUDE_PATH
python setup.py build_ext develop

🎯Running code

1) Configuration file: change the dataset root and the model save root by yourself

cd configs/evisseg_evuav.yaml

2) Training

train.py

3) Testing

test.py

4) Pre_trained weights

The pre_trained weights can be download here. Extracted code:ztp4

---

Citation

If you use this work in your research, please cite it:

@misc{chen2025eventbasedtinyobjectdetection,
      title={Event-based Tiny Object Detection: A Benchmark Dataset and Baseline}, 
      author={Nuo Chen and Chao Xiao and Yimian Dai and Shiman He and Miao Li and Wei An},
      year={2025},
      eprint={2506.23575},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.23575}, 
}

---

Acknowledgement

The code is based on HAIS and spconv.