Abstract: Virtual reality(VR), a new type of simulation and interaction technology, has arousedwidespread attention and research interest. It is necessary to evaluate thevirtual reality quality and provide a standard for the rapidly developingtechnology. To the best of our knowledge, few researchers have built benchmarkdatabases and designed related algorithms, which has hindered the furtherdevelopment of VR technology. In this work, a free available dataset (VRQ-TJU)for virtual reality quality assessment is presented with subjective scores foreach sample data. The validity for the designed database has been proved basedon the traditional multimedia quality assessment metrics. In addition, anend-to-end 3D convolutionalneuralnetworks (CNN) is introduced to predict the VR video quality without areferenced VR video. This method can extract spatio-temporal features and doesnot require using hand-crafted features. At the same time, a new score fusionstrategy is designed, based on the characteristics of the VR video projectionprocess. Taking the pre-processed VR video patches as input, the networkcaptures local spatiotemporal features and gets the score of every patch. Thenthe new quality score fusion strategy is applied to get the final score. Suchapproach shows advanced performance on this database.
Keywords: BenchmarkDatabase, Quality Score Fusion Strategy, Spatio-temporal Features, VirtualReality Quality Assessment, 3D Convolutional Neural Networks.