EventSegNet: Direct Sparse Semantic Segmentation from Event Data DOI Creative Commons
Pengju Li, Yuqiang Fang, Jiayu Qiu

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 17(1), P. 84 - 84

Published: Dec. 29, 2024

Semantic segmentation tasks encompass various applications, such as autonomous driving, medical imaging, and robotics. Achieving accurate semantic information retrieval under conditions of high dynamic range rapid scene changes remains a significant challenge for image-based algorithms. This is primarily attributable to the limitations conventional image sensors, which can experience motion blur or exposure artifacts. In contrast, event-based vision asynchronously report in pixel intensity, offer compelling solution by acquiring visual at same rate dynamics, thereby mitigating these limitations. However, we encounter tasks: need expend time on converting event data into frame images align with existing techniques. approach squanders inherently temporal resolution data, compromising accuracy real-time performance tasks. To address issues, this work explores sparse that directly addresses data. We propose network named EventSegNet improves ability extract geometric features from combining feature enhancement operations attention mechanisms. Based this, large-scale dataset provides labels each event. Our achieved new F1 score 84.2% dataset. addition, lightweight edge-oriented AI inference deployment technique was implemented model. Compared baseline model, optimized model reduces 1.1% but more than twice fast computationally, enabling NVIDIA AGX Xavier.

Language: Английский

Neuromorphic event-based recognition boosted by motion-aware learning DOI
Yuhan Liu, Yongjian Deng, Bochen Xie

et al.

Neurocomputing, Journal Year: 2025, Volume and Issue: unknown, P. 129678 - 129678

Published: Feb. 1, 2025

Language: Английский

Citations

0
ASOD-YOLOX: a study on small object detection in aerial images based on YOLOX DOI
H. Zhang, Wentao Liu,

Enyao Chen

et al.

The Journal of Supercomputing, Journal Year: 2025, Volume and Issue: 81(5)

Published: April 16, 2025

Language: Английский

Citations

0
LCFF-Net: A lightweight cross-scale feature fusion network for tiny target detection in UAV aerial imagery DOI Creative Commons
D. Tang,

Shuyun Tang,

Zhipeng Fan

et al.

PLoS ONE, Journal Year: 2024, Volume and Issue: 19(12), P. e0315267 - e0315267

Published: Dec. 19, 2024

In the field of UAV aerial image processing, ensuring accurate detection tiny targets is essential. Current target algorithms face challenges such as low computational demands, high accuracy, and fast speeds. To address these issues, we propose an improved, lightweight algorithm: LCFF-Net. First, LFERELAN module, designed to enhance extraction features optimize use resources. Second, a cross-scale feature pyramid network (LC-FPN) employed further enrich information, integrate multi-level maps, provide more comprehensive semantic information. Finally, increase model training speed achieve greater efficiency, lightweight, detail-enhanced, shared convolution head (LDSCD-Head) original head. Moreover, present different scale versions LCFF-Net algorithm suit various deployment environments. Empirical assessments conducted on VisDrone dataset validate efficacy proposed. Compared baseline-s model, LCFF-Net-n outperforms by achieving 2.8% in mAP 50 metric 3.9% improvement 50–95 metric, while reducing parameters 89.7%, FLOPs 50.5%, computation delay 24.7%. Thus, offers accuracy speeds for images, providing effective solution.

Language: Английский

Citations

0
EventSegNet: Direct Sparse Semantic Segmentation from Event Data DOI Creative Commons
Pengju Li, Yuqiang Fang, Jiayu Qiu

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 17(1), P. 84 - 84

Published: Dec. 29, 2024

Semantic segmentation tasks encompass various applications, such as autonomous driving, medical imaging, and robotics. Achieving accurate semantic information retrieval under conditions of high dynamic range rapid scene changes remains a significant challenge for image-based algorithms. This is primarily attributable to the limitations conventional image sensors, which can experience motion blur or exposure artifacts. In contrast, event-based vision asynchronously report in pixel intensity, offer compelling solution by acquiring visual at same rate dynamics, thereby mitigating these limitations. However, we encounter tasks: need expend time on converting event data into frame images align with existing techniques. approach squanders inherently temporal resolution data, compromising accuracy real-time performance tasks. To address issues, this work explores sparse that directly addresses data. We propose network named EventSegNet improves ability extract geometric features from combining feature enhancement operations attention mechanisms. Based this, large-scale dataset provides labels each event. Our achieved new F1 score 84.2% dataset. addition, lightweight edge-oriented AI inference deployment technique was implemented model. Compared baseline model, optimized model reduces 1.1% but more than twice fast computationally, enabling NVIDIA AGX Xavier.

Language: Английский

Citations

0