计算机科学
采样(信号处理)
推论
光流
理论(学习稳定性)
卷积神经网络
时间分辨率
还原(数学)
人工智能
模式识别(心理学)
算法
时态数据库
运动估计
卷积(计算机科学)
事件(粒子物理)
人工神经网络
数据建模
点过程
区间(图论)
状态空间
数据挖掘
国家(计算机科学)
点(几何)
时间尺度
循环神经网络
迭代重建
空间分析
重要性抽样
估计理论
计算机视觉
作者
Xize Cheng,Jianming Wang,Yukuan Sun,Xuesong Li,Jing Liu,Xiaojie Duan
标识
DOI:10.1117/1.jei.35.4.043002
摘要
Event-based optical-flow estimation remains challenging across different fixed-frequency sampling, particularly under low-frequency conditions where observations are sparse and motion displacements are large. Conventional methods based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) update their states at fixed discrete-time steps, ignoring the actual temporal intervals between events, which can lead to inconsistent temporal modeling and reduced accuracy. To address these challenges, we propose TSR-SSM, a temporal scale-robust state-space model for event-based optical-flow estimation. TSR-SSM discretizes the state space according to the temporal interval of each fixed sampling setting, enabling consistent temporal modeling under each fixed sampling setting. In addition, we design an SSM-enhanced cascaded decoder that integrates spatial state-space modeling with stability regularization, capturing long-range spatial dependencies while suppressing high-frequency noise. Experiments on the HREM and DSEC datasets show that TSR-SSM consistently outperforms prior methods, achieving a 4.2% reduction in average endpoint error under standard sampling and a 9.7% reduction under low-frequency conditions while maintaining real-time inference speed at 37.46 FPS on DSEC. These results demonstrate that TSR-SSM enables accurate, robust, and efficient event-based optical-flow estimation across diverse temporal scales.
科研通智能强力驱动
Strongly Powered by AbleSci AI