No-Reference 360° Video Quality Measurement With Local and Global Features Fusion

The no-reference (NR) 360° video quality measurement (VQM) method can measure the perceived quality of 360° video without using references, which is highly desired in immersive media applications such as virtual reality (VR). However, due to the lack of references, the performance of the method is usually unsatisfactory. In this article, to perform high efficiency NR 360° VQM, we propose to generate pseudo-references from distorted 360° videos for quality-aware local spatial feature extraction. To avoid projection distortion in 360° videos, this feature is extracted from the salient viewport. Additionally, we extract global spatial features from the entire 360° video frame utilizing the spherical convolutional neural network (CNN), and temporal features from video motion utilizing the fast paths of the SlowFast network to measure the 360° video quality from different perspectives. By concatenating and fusing these efficient features, the proposed method can measure 360° video quality with high accuracy. The proposed method has been compared favorably with the state-of-the-art methods to demonstrate its excellence, and the efficiency of different modules in the proposed method has also been well analyzed. The code is available at https://github.com/yeonhw/360BVQM.