A Consideration on Applying Fine-tuning to Subjective Evaluation Estimation of Videos to Reduce Subjects’ Loads

In subjective evaluation estimation of viewed videos using subjects’ bio-signals, increasing the number of showing videos has been reported to be effective in improving estimation accuracy. However, it increases the subjects’ loads and fatigue that prevent accurate bio-signal acquisition. In this paper, thus, we propose to apply fine-tuning to a pretrained model constructed by using other subjects’ data who have similar bio-signals of the target subject to decrease the number of videos watched by that subject. To verify the proposed method, a public dataset of DEAP is used for our experiment, which shows almost equal or better estimation accuracy can be achieved by using a smaller number of videos in the four categories of subjective evaluation.