在视觉监控系统中,有必要认识到人们处理诸如电话,杯子或塑料袋之类物体的行为。在本文中,为了解决这个问题,我们提出了一个新的框架,用于通过图形卷积网络使用人类和对象姿势识别与对象相关的人类行为。在此框架中,我们通过选择性地对视频中的信息帧进行采样来构建可靠人类的骨架图,其中包括在姿势估计中获得的具有高置信度分数的人类关节。从采样帧生成的骨架图表示与空间域和时域中的对象位置相关的人体姿势,并且这些图被用作图卷积网络的输入。通过开放基准和我们自己的数据集进行实验,我们验证了框架的有效性,因为我们的方法优于基于骨架的动作识别的最先进方法。
translated by 谷歌翻译
在人员重新识别(ReID)任务中,由于其缺乏可训练的数据集,通常使用在大型数据集上预训练的分类网络的微调方法。然而,由于梯度问题,相对难以有效地微调网络的低级层。在这项工作中,我们提出了一种新颖的微调策略,它允许通过将高级层的权重回滚到其初始预训练权重来充分训练低级层。我们的战略目标是缓解低级别层中梯度消失的问题,并且可以有效地跟踪低级别层以适应ReID数据集,从而提高ReID任务的性能。通过几个实验验证了所提出的策略的改进性能。此外,如果没有任何附加功能,例如估计或分段,我们的策略只使用vanilla深度卷积神经网络架构展示最先进的性能。
translated by 谷歌翻译
神经元的激活边界是指分离超平面,其确定神经元是激活还是停用。在神经网络中长期以来认为,神经元的激活,而不是精确的输出值,在隐藏特征空间的形成分类友好分区中起着最重要的作用。然而,正如我们所知道的那样,神经网络的这个方面在知识转移的文献中没有被考虑过。在本文中,我们通过蒸馏由隐藏神经元形成的激活边界提出了一种知识转移方法。对于蒸馏,我们提出激活转移损失,当学生产生的边界与教师一致时,该转移损失具有最小值。由于激活转移损失不可微分,我们设计了近似于激活转移损失的分段可微分损失。通过该方法,学生学习了教师中每个神经元形成的激活区域和去激活区域之间的分离边界。通过知识转移的各个方面的实验,验证了所提出的方法优于当前的最新技术水平。 。
translated by 谷歌翻译
Many recent works on knowledge distillation have provided ways to transfer the knowledge of a trained network for improving the learning process of a new one, but finding a good technique for knowledge distillation is still an open problem. In this paper, we provide a new perspective based on a decision boundary, which is one of the most important component of a classifier. The generalization performance of a classifier is closely related to the adequacy of its decision boundary, so a good classifier bears a good decision boundary. Therefore , transferring information closely related to the decision boundary can be a good attempt for knowledge distillation. To realize this goal, we utilize an adversarial attack to discover samples supporting a decision boundary. Based on this idea, to transfer more accurate information about the decision boundary, the proposed algorithm trains a student classi-fier based on the adversarial samples supporting the decision boundary. Experiments show that the proposed method indeed improves knowledge distillation and achieves the state-of-the-arts performance. 1
translated by 谷歌翻译
基于深度学习的图像分割方法的最新进展已经具有人类精确度的实时性能。然而,偶尔最好的方法由于低图像质量,伪像或黑盒算法的意外行为而失败。能够在没有基础事实的情况下预测分割质量在临床实践中是至关重要的,但在大规模研究中也是如此,以避免在随后的分析中包含无效数据。在这项工作中,我们提出了两种使用深度学习进行心血管MR分割的实时自动质量控制方法。首先,在12,880个样本上对一个神经网络进行润湿,以便根据每个案例预测Dice相似系数(DSC)。我们报告1,610个测试样本的平均误差(MAE)为0.03,二元分类精度为97%,这反映了低质量和高质量的分割。其次,在没有手动注释数据可用的情况下,我们训练网络来预测通过反向测试策略获得的估计质量的DSC分数。对于这种情况,Wereport的MAE = 0.14和91%二进制分类精度。实时获得预测,当与实时分割方法结合时,能够在患者仍在扫描仪中时获得关于获取的扫描是否是可分析的即时反馈。这进一步使得优化图像采集的新应用朝向最佳可能的分析结果。
translated by 谷歌翻译
MR成像将在肿瘤体积和器官分割的放射治疗计划中发挥非常重要的作用。然而,由于高成本和在老龄化社会中增加使用诸如心脏起搏器和人工关节的金属植入物,基于MR的放射疗法的使用是有限的。为了提高基于CT的放射治疗计划的准确性,我们提出了一种合成方法,使用配对和非配对训练数据将CT图像转换为MR图像。与当前用于医学图像的合成方法(其依赖于稀疏成对比对数据或大量未配对数据)相比,所提出的方法减轻了配对训练的刚性配准挑战并且克服了未配对训练的上下文未对准问题。训练生成对抗网络以将2D脑CTimage切片转换为2D脑MR图像切片,结合对抗性损失,双周期一致性损失和体素消失。使用202名患者的CT和MR图像分析实验。针对独立配对训练和不成对训练方法的定性和定量比较证明了我们的方法的优越性。
translated by 谷歌翻译
我们最近看到许多成功应用复发神经网络(RNN)的电子病历(EMR),其中包含患者的诊断,药物和其他各种事件的历史,以便预测患者的当前和未来状态。尽管RNN具有强大的性能,但用户通常很难理解为什么模型会做出特定的预测。 RNN的这种黑盒性质可能阻碍临床实践的广泛采用。此外,我们没有确定的方法来交互式地利用用户的领域专业知识和先前的知识输入来指导模型。因此,我们的设计研究旨在通过医学专家,人工智能科学家和视觉分析研究人员的共同努力,提供可视化分析解决方案,以提高NRN的可解释性和互动性。在专家之间的迭代设计过程之后,我们设计,实施和评估了一个名为RetainVis的可视化分析工具,它结合了一个新改进的,可解释的和交互式的基于RNN的模型RetainEX和可视化,供用户在预测任务环境中探索EMR数据。我们的研究表明,使用RetainVis可以有效利用个体医疗代码如何利用心力衰竭和白内障症状患者的EMR进行风险预测。我们的研究还展示了我们如何对名为RETAIN的最先进的RNN模型进行实质性改变,以便利用时间信息并增加交互性。本研究将为研究人员提供有用的指导方针,旨在为RNN设计可解释和交互式的可视化分析工具。
translated by 谷歌翻译
Research indicates that human-robot interaction can help children with autism spectrum disorder (ASD). While most early robot-mediated interaction studies were based on free interactions , recent studies have shown that robot-mediated interventions that focus on the core impairments of ASD such as joint attention deficit tend to produce better outcomes. Joint attention impairment is one of the core deficits in ASD that has an important impact in the neuropsychological development of these children. In this paper, we propose a novel joint attention intervention system for children with ASD that overcomes several existing limitations in this domain such as the need to use body-worn sensors, nonau-tonomous robot operation requiring human involvement and lack of a formal model for robot-mediated joint attention interaction. We present a fully autonomous robotic system, called noncontact-responsive robot-mediated intervention system, that can infer attention through a distributed noncontact gaze inference mechanism with an embedded least-to-most (LTM) robot-mediated interaction model to address the current limitations. The system was tested in a multisession user study with 14 young children with ASD. The results showed that participants' joint attention skills improved significantly, their interest in the robot remained consistent throughout the sessions, and the LTM interaction model was effective in promoting the children's performance.
translated by 谷歌翻译
Our research provides social scientists with areas of inquiry in tobacco-related health disparities in young adult women and opportunities for intervention, as Instagram may be a powerful tool for the public health surveillance of smoking behavior and social norms among young women. Social media has fundamentally changed how to engage with health-related information. Researchers increasingly turn to social media platforms for public health surveillance. Instagram currently is one of the fastest growing social networks with over 53% of young adults (aged 18-29) using the platform and young adult women comprise a significant user base. We conducted a content analysis of a sample of smoking imagery drawn from Instagram's public Application Programming Interface (API). From August 2014 to July 2015, 18 popular tobacco-and e-cigarette-related text tags were used to collect 2.3 million image posts. Trained undergraduate coders (aged 21-29) coded 8,000 images (r = .91) by type of artifact, branding, number of persons, gender, age, ethnicity, and the presence of smoke. Approximately 71.5% of images were tobacco-relevant and informed our research. Images of cigarettes were the most popular (49%), followed by e-cigarettes (32.1%). "Selfies while smoking" was the dominant form of portrait expression, with 61.4% of images containing only one person, and of those, 65.7% contained images of women. The most common selfie was women engaged in "smoke play" (62.4%) that the viewer could interpret as "cool." These "cool" images may counteract public health efforts to denormalize smoking, and young women are bearing the brunt of this under-the-radar tobacco advertising. Social media further normalizes tobacco use because positive images and brand messaging are easily seen and shared, and also operates as unpaid advertising on image-based platforms like Instagram. These findings portend a dangerous trend for young women in the absence of effective public health intervention strategies.
translated by 谷歌翻译
² Aging is a process that is inevitable, and makes our body vulnerable to age-related diseases. Age is the most consistent factor affecting the sleep structure. Therefore, new automatic sleep staging methods, to be used in both of young and elderly patients, are needed. This study proposes an automatic sleep stage detector, which can separate wakefulness, rapid-eye-movement (REM) sleep and non-REM (NREM) sleep using only EEG and EOG. Most sleep events, which define the sleep stages, are reduced with age. This is addressed by focusing on the amplitude of the clinical EEG bands, and not the affected sleep events. The age-related influences are then reduced by robust subject-specific scaling. The classification of the three sleep stages are achieved by a multi-class support vector machine using the one-versus-rest scheme. It was possible to obtain a high classification accuracy of 0.91. Validation of the sleep stage detector in other sleep disorders, such as apnea and narcolepsy, should be considered in future work.
translated by 谷歌翻译