基于人机交互的深度学习训练数据标注系统

Human-computer Interaction-based Deep Learning and Training Data Annotation System

目前的数据标注平台以及开源数据标注工具普遍存在多人合作的标注流程不合理的问题,无法保证标注的效率和质量。针对该问题,提出一种结对标注法,采用两两分组,同时标注,互相审查的方式进行标注。实验证明,结对标注法可以提高63%的标注效率。另外,提出推测标注法,当输入数据为视频时,基于数据之间的联系,使标注工作量降低为未推测标注的一半。实验证明,推测标注法可以提高25%标注效率。

The current data labeling platforms and open-source data labeling tools generally have the problem of unreasonable labeling process with multi-people cooperation,which cannot guarantee the effi ciency and quality of labeling.To address this problem,this paper proposes a pair annotation method,which uses a way of pair groups,simultaneous annotation and review of each other to annotate.The experiments prove that the pair annotation method can significantly improve the annotation effi ciency by 63%.In addition,this paper proposes a speculative annotation method,which reduces the annotation workload to half of the unprojected annotation based on the connection between the data when the input data is video.It is demonstrated that the speculative annotation method can signifi cantly improve the annotation effi ciency by 25%.