VR sickness是阻礙VR市場發展的障礙之一。因此,VR sickness的客觀指標非常重要,可以避免使用者遭受VR sickness。最近,一些研究利用深度學習方法來預測VR sickness。但是,他們的方法需要昂貴的計算資源,這限制了real-time的應用。在本文中,我們提出了從光流導出的混合時間特徵,即horizontal motion strength、vertical motion strength和motion anisotropy。我們建立新的數據集,其中包含二十部5分鐘長的360度影片。在實驗中,每位受試者每分鐘會回答Discomfort Scores(0-10),並在影片結束時填寫模擬器疾病問卷(SSQ)。最後,我們採用隨機森林模型在數據集上進行訓練。該模型使用當前混合時間特徵和先前時段的混合時間特徵,這使得模型考慮了VR sickness之間的時間依賴性。我們的方法分別在PLCC和SROCC上比state-of-the-art高出2%和4%,並且可以real-time運行。
VR sickness is one of the obstacles hindering the development of the VR market. Therefore, the objective metric of VR sickness is very important that can help the user to avoid suffering VR sickness. Recently, some works utilize deep learning methods to predict VR sickness. However, their methods are computationally expensive which limited applications in real-time tasks. In this paper, hybrid temporal features derived from optical flow are proposed, namely horizontal motion strength, vertical motion strength, and motion anisotropy. We introduce a new dataset that contains twenty 5-minute-long 360-degree videos. In the experiment, each subject answers Discomfort Scores (0-10) every minute and performs the Simulator Sickness Questionnaire (SSQ) scores at the end of the video. Finally, a random forest model is adopted to train on our dataset. The model uses the current hybrid temporal features and the previous time period hybrid temporal features that make the model considers temporal dependency between the degree of VR sickness. This method not only outperformed previous state-of-the-art by 2% on PLCC and 4% on SROCC but also runs in real-time.