基于视觉特征与对抗学习的大五人格评估方法

Big Five Personality Assessment Method Based on Visual Featuresand Adversarial Learning

由于传统的依托于人格量表的大五人格评估方法效率极低,而现有的基于多模态的人格评估方法在实用中很难满足多通道的高质量输入需求。针对这种情况,提出了一种基于纯粹的视觉特征的人格评估方法,并与对抗学习相结合,通过一个人日常生活中的视频,获取其大五人格的定量评估。首先,通过对视频图片帧进行建模,并结合对抗学习的思想,使特征提取器尽力提取与性别无关,而与人格有关的特征。其次,利用多粗细粒度损失结构的注视估计模型,获取人物注视分布及其视线序列特征,最后对提取的所有视觉特征进行融合,并对其大五人格进行预测。实验结果表明,该方法的平均预测精度为91.96%,在ECCV 2016年人格挑战数据集中名列前茅,并且超过了此前所有的基于视觉特征的大五人格评估方法。

Due to the low efficiency of the traditional big five personality assessment method based on personality scale,the existing personality assessment method based on multimodality is difficult to meet the demand of multi-channel high-quality input in practice.In view of this situation,a personality assessment method based on pure visual features is proposed,which is combined with adversarial learning to obtain the quantitative assessment of a person's major five personality through the video in his daily life.First,by modeling video frames and combining the idea of adversarial learning,the feature extractor tries its best to extract features that are not related to gender but related to personality.Secondly,the gaze estimation model with multi coarse and fine grain loss structure is used to obtain the gaze distribution and gaze sequence features of the characters.Finally,all the extracted visual fea⁃tures are fused and their major five personalities are predicted.The experimental results show that the average prediction accuracy of this method is 91.96%,which is among the best in the ECCV personality challenge data set in 2016,and exceeds all previous big five personality assessment methods based on visual features.