Saliency Supervision for Pain Intensity Regression

EasyChair Preprint 511

Abstract

Obtaining pain intensity values from face images is an important problem in autonomous nursing systems. Although deep neural networks have achieved great success in many fields of computer vision, adopting deep neural networks is difficult due to the limited data and the subjectiveness of obtaining pain intensity values. Inspired by the prior of human vision systems, we propose a novel approach called saliency supervision, where the deep neural networks are directly regularized to focus on the facial area which is discriminative for pain regression. Through alternative training between saliency supervision and global loss, our method can learn sparse and robust features which is proved helpful for pain intensity regression. We verify our method with the face verification network backbone on the UNBC-McMaster Shoulder-Pain dataset, and achieve state-of-art performance without bells and whistles.

Keyphrases: Regression, multi-task training, regularization, saliency supervision, triplet loss

@booklet{EasyChair:511,
  author    = {Conghui Li and Zhaocheng Zhu and Yuming Zhao},
  title     = {Saliency Supervision for Pain Intensity Regression},
  doi       = {10.29007/67lb},
  howpublished = {EasyChair Preprint 511},
  year      = {EasyChair, 2018}}