N-DHF1K; N-UCF Sports
收藏资源简介:
本研究构建了N-DHF1K与N-UCF Sports两个合成事件显著性数据集,旨在解决事件相机领域缺乏大规模标注数据的瓶颈。N-DHF1K源自DHF1K视频显著性基准,包含1000个涵盖150种场景类别的视频,共超过60万帧,并融合了17位观察者的逐帧注视点标注;N-UCF Sports则基于UCF Sports动作数据集,包含150段涵盖9类体育动作的视频。数据集通过ESIM库将RGB视频流转化为事件流,采用正负对比度阈值0.09与3毫秒不应期参数,完整保留了运动、对比度与时间显著性线索。这些合成数据为事件驱动的视觉注意力建模提供了规模化监督信号,支持动态显著性预测模型的训练与评估,推动神经形态视觉与注意力机制交叉领域的发展。
This study constructed two synthetic event saliency datasets, N-DHF1K and N-UCF Sports, to address the bottleneck of lacking large-scale labeled data in the field of event cameras. N-DHF1K is derived from the DHF1K video saliency benchmark, containing 1000 videos covering 150 scene categories, with over 600,000 frames in total, and incorporating frame-by-frame gaze annotations from 17 observers. N-UCF Sports is based on the UCF Sports action dataset, which includes 150 videos covering 9 categories of sports actions. The datasets converted RGB video streams into event streams via the ESIM library, adopting positive and negative contrast thresholds of 0.09 and a 3-millisecond refractory period parameter, which fully retained motion, contrast, and temporal saliency cues. These synthetic datasets provide large-scale supervision signals for event-driven visual attention modeling, supporting the training and evaluation of dynamic saliency prediction models, and promoting the development of the interdisciplinary field of neuromorphic vision and attention mechanisms.
该数据集详情页面目前尚未提供具体的数据集信息。页面中仅标注“代码和数据将很快发布”,目前无任何可供总结的数据集内容、结构、用途或示例。

- 1Exploring deep learning for Event-Based Saliency Prediction with a Transformer-based model法国国家科学研究中心; 蔚蓝海岸大学 · 2026年




