STGCN360: A QoE Assessment Method with Spatio-Temporal Graph Convolutional Networks for Tile-Based 360° Video Streaming

With the proliferation of 5G, wireless networks, and other infrastructure, 360° video streaming has experienced rapid development. Efficient scheduling of 360° video streams relies on accurate feedback of user-side Quality of Experience (QoE), necessitating the construction of more precise QoE assessment methods. However, the existing tile-based QoE assessment methods for 360° video streaming have several limitations. First, full-reference video quality assessment methods require additional overhead to transmit the source video as a reference. Second, many learning-based methods are too complex. This high complexity results in heavy computational costs during training, reducing their practicality in scenarios requiring frequent model adaptation. Third, some methods rely only on simple indicators like bitrate and stall duration or spatial features in isolation. They ignore the spatio-temporal coupling inherent in 360° videos, which reduces the QoE assessment accuracy. To sum up, there is a lack of a lightweight QoE assessment method that can effectively integrate multidimensional influencing factors like the spatio-temporal features of 360° video and network state. A matching QoE assessment dataset is also missing. Therefore, focusing on tile-based 360° video streaming, this paper proposes a QoE assessment method named STGCN360. This method comprehensively considers multidimensional influencing factors, including network state and the spatio-temporal features of the video stream. To reduce complexity, it limits spatio-temporal graph modeling to the key tiles within the user’s viewport, avoiding the need to process all tiles. Then, a spatio-temporal graph convolutional network (STGCN) is employed to train the QoE assessment model. Furthermore, we integrate multi-source heterogeneous datasets through feature engineering, enabling the simultaneous representation of both video quality and multidimensional factors to support the training of STGCN360. The experimental results indicate that, compared to the existing methods, STGCN360 enables more accurate QoE assessment for 360° video streaming, improving accuracy by approximately 30.79% to 32.07%. Simultaneously, the training time cost is significantly reduced, with training efficiency improved by approximately 3.7 to 5.1 times.