人工智能培训

搜索

人工智能论文:用于高效视频推理的在线模型蒸馏(Online Model Distillation for Efficient Video Inference)

[复制链接]
kjdshfjsfgsfg 发表于 2018-12-7 10:37:11 | 显示全部楼层 |阅读模式
kjdshfjsfgsfg 2018-12-7 10:37:11 178 0 显示全部楼层
人工智能论文:用于高效视频推理的在线模型蒸馏(Online Model Distillation for Efficient Video Inference)高质量的计算机视觉模型通常解决了解真实世界图像的一般分布的问题。然而,大多数相机只观察到这种分布的很小一部分。这提供了通过将紧凑的低成本模型专门用于由单面板观察到的特定分布框架来实现更有效推断的可能性。在本文中,我们采用模型蒸馏技术(使用高成本教师的输出监督低成本学生模型),将精确,低成本的语义分割模型专门化为目标视频流。我们不是从视频流中学习离线数据的专业学生模型,而是通过实时视频在线培训学生,间歇性地运行教师以提供学习目标。 Onlinemodel蒸馏产生语义分割模型,即使目标视频的分布是非静态的,它们也会使Mask R-CNN教师接近7到17倍的推理运行时成本(11到26x FLOP)。我们的方法不需要对目标视频流进行离线预训练,并且比基于流或视频对象分割的解决方案实现更高的准确性和更低的成本。我们还提供了一个新的视频数据集,用于评估长时间运行的视频流的推理效率。
High-quality computer vision models typically address the problem ofunderstanding the general distribution of real-world images.However, mostcameras observe only a very small fraction of this distribution.This offersthe possibility of achieving more efficient inference by specializing compact,low-cost models to the specific distribution of frames observed by a singlecamera.In this paper, we employ the technique of model distillation(supervising a low-cost student model using the output of a high-cost teacher)to specialize accurate, low-cost semantic segmentation models to a target videostream.Rather than learn a specialized student model on offline data from thevideo stream, we train the student in an online fashion on the live video,intermittently running the teacher to provide a target for learning.Onlinemodel distillation yields semantic segmentation models that closely approximatetheir Mask R-CNN teacher with 7 to 17x lower inference runtime cost (11 to 26xin FLOPs), even when the target video's distribution is non-stationary.Ourmethod requires no offline pretraining on the target video stream, and achieveshigher accuracy and lower cost than solutions based on flow or video objectsegmentation.We also provide a new video dataset for evaluating the efficiencyof inference over long running video streams.人工智能论文:用于高效视频推理的在线模型蒸馏(Online Model Distillation for Efficient Video Inference) mrg8aBAgd37Dlrb6.jpg
URL地址:https://arxiv.org/abs/1812.02699     ----pdf下载地址:https://arxiv.org/pdf/1812.02699    ----人工智能论文:用于高效视频推理的在线模型蒸馏(Online Model Distillation for Efficient Video Inference)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

kjdshfjsfgsfg当前离线
新手上路

查看:178 | 回复:0

快速回复 返回顶部 返回列表