目的视频动作质量评估旨在评估视频中特定动作的执行情况和完成质量。自动化的动作质量评估能够有效地减少人力资源的损耗,可以更加精准、公正地对视频内容进行评估。传统动作质量评估方法主要存在以下问题:(1)视频中动作主体的多尺度...目的视频动作质量评估旨在评估视频中特定动作的执行情况和完成质量。自动化的动作质量评估能够有效地减少人力资源的损耗,可以更加精准、公正地对视频内容进行评估。传统动作质量评估方法主要存在以下问题:(1)视频中动作主体的多尺度时空特征问题;(2)认知差异导致的标记内在模糊性问题;(3)多头自注意力机制的注意力头冗余问题。针对以上问题,提出了一种能够感知视频序列中不同时空位置、生成细粒度标记的动作质量评估模型SALDL(self-attention and label distribution learning)。方法SALDL提出Attention-Inc(attention-inception)结构,该结构通过Embedding、多头自注意力以及多层感知机将自注意力机制渐进式融入Inception结构,使模型能够获得不同尺度卷积特征之间的上下文信息。提出一种正负时间注意力模块PNTA(pos-neg temporal attention),通过PNTA损失挖掘时间注意力特征,从而减少自注意力头冗余并提取不同片段的注意力特征。SALDL模型通过标记增强及标记分布学习生成细粒度的动作质量标记。结果提出的SALDL模型在MTL-AQA(multitask learning-action quality assessment)和JIGSAWS(JHU-ISI gesture and skill assessment working set)等数据集上进行了大量对比及消融实验,斯皮尔曼等级相关系数分别为0.9416和0.8183。结论SALDL模型通过充分挖掘不同尺度的时空特征解决了多尺度时空特征问题,并引入符合标记分布的先验知识进行标记增强,达到了解决标记的内在模糊性问题以及注意力头的冗余问题。展开更多
Exposure to fine particulate matter(PM2.5)is known to harm public health.In China,after implementation of aggressive emissions control measures under the Action Plan of Air Pollution Prevention and Control(2013-2017),...Exposure to fine particulate matter(PM2.5)is known to harm public health.In China,after implementation of aggressive emissions control measures under the Action Plan of Air Pollution Prevention and Control(2013-2017),air quality has significantly improved.In this work,we investigated changes in PM2.5 exposure and the associated health impacts in China for the period 2013-2017.We used an optimal estimator of PM2.5 combining in-situ observations,satellite measurements,and simulations from a chemical transport model to derive the spatial and temporal variations in PM2.5 exposure,and then used welldeveloped exposure-response functions to estimate the premature deaths attributable to PM2.5 exposure.We found that national population-weighed annual mean PM2.5 concentrations decreased from 67.4μgm-3 in 2013 to 45.5μgm-3 in 2017(32%reduction).This rapid decrease in PM2.5 pollution led to a 14%reduction in premature deaths due to long-term exposure.We estimated that,during 2013-2017,the premature deaths attributable to long-term PM2.5 exposure decreased from 1.2 million(95%CI:1.0,1.3;fraction of total mortality:13%)in 2013 to 1.0 million(95%CI:0.9,1.2;10%)in 2017.Despite the rapid decrease in annual mean PM2.5 concentrations,health benefits associated with reduced long-term exposure were limited,because for many cities,the PM2.5 levels remain at the portion where the exposure-response function is less steeper than that at the lowconcentration end.We also found that the deaths associated with acute exposure decreased by 61%during 2013-2017 due to rapid reduction in the number of heavily polluted days.Our results confirm that clean air policies in China have mitigated the air pollution crisis;however,continuous emissions reduction efforts are required to protect citizens from air pollution.展开更多
文摘目的视频动作质量评估旨在评估视频中特定动作的执行情况和完成质量。自动化的动作质量评估能够有效地减少人力资源的损耗,可以更加精准、公正地对视频内容进行评估。传统动作质量评估方法主要存在以下问题:(1)视频中动作主体的多尺度时空特征问题;(2)认知差异导致的标记内在模糊性问题;(3)多头自注意力机制的注意力头冗余问题。针对以上问题,提出了一种能够感知视频序列中不同时空位置、生成细粒度标记的动作质量评估模型SALDL(self-attention and label distribution learning)。方法SALDL提出Attention-Inc(attention-inception)结构,该结构通过Embedding、多头自注意力以及多层感知机将自注意力机制渐进式融入Inception结构,使模型能够获得不同尺度卷积特征之间的上下文信息。提出一种正负时间注意力模块PNTA(pos-neg temporal attention),通过PNTA损失挖掘时间注意力特征,从而减少自注意力头冗余并提取不同片段的注意力特征。SALDL模型通过标记增强及标记分布学习生成细粒度的动作质量标记。结果提出的SALDL模型在MTL-AQA(multitask learning-action quality assessment)和JIGSAWS(JHU-ISI gesture and skill assessment working set)等数据集上进行了大量对比及消融实验,斯皮尔曼等级相关系数分别为0.9416和0.8183。结论SALDL模型通过充分挖掘不同尺度的时空特征解决了多尺度时空特征问题,并引入符合标记分布的先验知识进行标记增强,达到了解决标记的内在模糊性问题以及注意力头的冗余问题。
基金supported by the National Natural Science Foundation of China (Grant Nos. 41571130032, 41571130035, 41625020 & 41701591)the National Key R & D Program (Grant No. 2016YFC0201506)
文摘Exposure to fine particulate matter(PM2.5)is known to harm public health.In China,after implementation of aggressive emissions control measures under the Action Plan of Air Pollution Prevention and Control(2013-2017),air quality has significantly improved.In this work,we investigated changes in PM2.5 exposure and the associated health impacts in China for the period 2013-2017.We used an optimal estimator of PM2.5 combining in-situ observations,satellite measurements,and simulations from a chemical transport model to derive the spatial and temporal variations in PM2.5 exposure,and then used welldeveloped exposure-response functions to estimate the premature deaths attributable to PM2.5 exposure.We found that national population-weighed annual mean PM2.5 concentrations decreased from 67.4μgm-3 in 2013 to 45.5μgm-3 in 2017(32%reduction).This rapid decrease in PM2.5 pollution led to a 14%reduction in premature deaths due to long-term exposure.We estimated that,during 2013-2017,the premature deaths attributable to long-term PM2.5 exposure decreased from 1.2 million(95%CI:1.0,1.3;fraction of total mortality:13%)in 2013 to 1.0 million(95%CI:0.9,1.2;10%)in 2017.Despite the rapid decrease in annual mean PM2.5 concentrations,health benefits associated with reduced long-term exposure were limited,because for many cities,the PM2.5 levels remain at the portion where the exposure-response function is less steeper than that at the lowconcentration end.We also found that the deaths associated with acute exposure decreased by 61%during 2013-2017 due to rapid reduction in the number of heavily polluted days.Our results confirm that clean air policies in China have mitigated the air pollution crisis;however,continuous emissions reduction efforts are required to protect citizens from air pollution.
