文章摘要
基于改进LiteFlowNet3网络的轻量化光流估计方法
Lightweight Optical Flow Estimation Method Based on Improved LiteFlowNet3
投稿时间:2024-01-24  修订日期:2024-05-09
DOI:
中文关键词: 光流估计;轻量化;LiteFlowNet3  深度可分离卷积
英文关键词: optical flow estimation  lightweight  LiteFlowNet3  depth-separable convolution
基金项目:安徽省高校自然科学重点项目(2022AH050249)
作者单位邮编
方潜生* 安徽建筑大学 230022
张亮 安徽建筑大学 
颜普 安徽建筑大学 
徐朝阳 安徽建筑大学 
摘要点击次数: 314
全文下载次数: 0
中文摘要:
      现有的光流估计方法参数量大,计算耗时,难以满足实时性需求。为解决这一问题,提出一种基于改进LiteFlowNet3网络的轻量化光流估计方法。使用池化和深度可分离卷积替换LiteFlowNet3网络中的常规卷积层,大幅减少网络模型参数量。在损失函数中增加了光流梯度的损失,强调训练中对光流边界的监督,在不增加参数情况下提升性能。改进的LiteFlowNet3网络参数量仅为0.78M。实验结果表明,改进的LiteFlowNet3光流估计方法在Sintel数据集的Clean和Final序列上的端点误差分别为2.69和4.12,同时单次推理时间仅为25ms。性能优于其他的轻量级光流方法,具有较强的竞争力。
英文摘要:
      Existing optical flow estimation methods have large parameter counts and time-consuming computations, which are difficult to meet the real-time demand. To solve this problem, a lightweight optical flow estimation method based on improved LiteFlowNet3 is proposed. Pooling and depth-separable convolution are used to replace the conventional convolutional layers in LiteFlowNet3, significantly reduce the number of network model parameters. The loss of the optical flow gradient is added to the loss function to emphasize the supervision of the optical flow boundaries during training, improved the performance without increasing the parameters. The improved LiteFlowNet3 parameter count is only 0.78 M. Experimental results show that the improved LiteFlowNet3 optical flow estimation method has end-point errors of 2.69 and 4.12 on Clean and Final sequences of the Sintel dataset. Respectively, the single inference time is only 25 ms. The performance outperforms that of other lightweight optical flow methods and is highly competitive.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭

分享按钮