Mobile 360° Video QoE: Measurement, Emulation, and GAN-Driven Throughput Synthesis

Mobile 360° Video QoE: Empirical Analysis of 5G QoS Metrics Raza Ul Mustafa Loyola University New Orleans USA [email protected] Md Tariqul Islam Universidade Estadual de Campinas (UNICAMP) Brazil [email protected] Roi Dupart Loyola University New Orleans USA [email protected] Noman Ashraf Instituto Polit´ ecnico Nacional Mexico [email protected] Christian Esteve Rothenberg Universidade Estadual de Campinas (UNICAMP) Brazil cheste[email protected] Abstract—The increasing demand for mobile 360-degree video streaming drives the need for highly immersive and smooth user experiences. However, maintaining a smooth Quality of Experience (QoE) for 360-degree video is challenging due to strict network requirements, particularly on variable mobile networks. These challenges can lead to poor streaming quality and cyber/motion sickness within immersive settings. Addressing these issues requires understanding the impact of network conditions and scalable testing under realistic scenarios. For this purpose, this paper presents an empirical analysis correlating real-world 5G radio Quality of Service (QoS) metrics from three U.S. operators with YouTube 360-degree video QoE. Our findings include a valuable dataset linking 5G radio QoS to video QoE. Index Terms—360-degree video, 5G, QoS, QoE I. INTRODUCTION The dominance of mobile video streaming in global internet traffic has been a well-discussed topic in recent years [1]. With rapid advancements in networking and multimedia technology, immersive media consumption, specifically 360-degree online video streaming, has gained significant popularity and is most evident on popular mobile platforms, such as YouTube, in recent times [2]. To ensure a satisfactory immersive Quality of Experience (QoE) for 360-degree video, high bandwidth requirements (up to 80 times more than standard streaming) and low latency (less than 9 milliseconds) are necessary [3]. These requirements vary significantly from the Key Performance Indicators (KPIs) for traditional 2D streaming over mobile networks. The nature of 360-degree video streaming provides users with a truly omnidirectional view, allowing for interaction with the content. However, the higher KPIs needed to stream 360-degree present considerable challenges. Failing to meet these standards can lead to unpleasant experiences (e.g., cyber/motion sickness) [2], [4]. Mobile network environments are naturally dynamic and heterogeneous, characterized by variations in signal quality This work was supported by Ericsson Telecomunicac¸ ˜ oes LTDA, and by the S˜ ao Paulo Research Foundation (FAPESP), grant 2021/00199-8, CPE SMARTNESS. This work was also partially funded by Loyola University New Orleans, USA. and available bandwidth. Due to its stringent KPI requirements, this variability directly influences streaming performance, particularly for immersive 360-degree video [2], [5]. Consequently, understanding and mitigating the effects of mobile network variability on immersive video QoE is crucial for stakeholders, including network and content providers and researchers, to optimize and improve the immersive experience for mobile users. The motivation for this work arises from the need to analyze and comprehend fluctuating network-level Quality of Service (QoS) parameters and their relationship to user-perceived QoE, particularly under the diverse conditions characteristic of real-world 5G mobile networks. In this context, this work involves an empirical analysis of the correlation between network behavior and corresponding immersive video streaming performance, based on empirical data collection from modern mobile network operators. The main contribution of this paper is an extensive empirical analysis of the correlation between real-world mobile network QoS (particularly at the radio layer) and user-perceived QoE for YouTube 360-degree video. This analysis is based on an in-the-wild measurement campaign conducted across three U.S. mobile operators using commercial smartphones and the GNetTrack Pro tool. II. BACKGROUND AND RELATED WORK 360-Video Streaming and Network Dynamics. 360-degree video streaming, particularly on platforms like YouTube, has become a prominent research topic. Unlike conventional 2D video, it requires specialized capture (e.g., omnidirectional cameras), projection formats such as equirectangular mapping, and tile-based encoding to support high spatial resolutions (e.g., 2K/4K). It can be rendered either on a user’s smartphone or within a head-mounted display. Due to its immersive and viewport-aware nature, 360 video can demand significantly larger bandwidth and ultra-low motion-to-photon latency to avoid cybersickness and ensure user immersion. Delivering this experience often relies on adaptive HTTP-based streaming (HAS) protocols, such as MPEG-DASH or HLS, enhanced with techniques like tiled streaming to optimize bandwidth by selectively delivering high-quality video segments corresponding to the user’s current viewport [4]. However, streaming 360-degree video over cellular or mobile networks presents challenges due to the higher network QoS requirements. Even in 5G networks, such variations in QoS directly affect the QoE, with issues like stalling, resolution drops, and delayed interaction responses leading to degraded immersion or physical discomfort [2], [5]. Therefore, understanding and adapting to mobile network dynamics is crucial for delivering high-quality 360-degree video experiences. Related Work. Prior research on 360-degree video has explored classification, QoE modeling, and network optimization. For instance, 360NorVic [6] proposed a machine-learning (ML) classifier to distinguish 360-degree video from regular encrypted traffic, achieving over 90% accuracy. In the context of QoE assessment, authors in [7] reviewed objective and subjective methods, highlighting the role of viewportaware and tile-based streaming strategies. Other works have investigated applying ML to XR/VR services: [8] introduced a framework for enhancing XR quality in 5G/6G networks, while [9], [10] applied ML and super-resolution techniques to improve 360 video quality under bandwidth limitations. Deep reinforcement learning has also been explored for adapting resource allocation in VR applications to dynamic wireless environments [11]. Compared to these studies, our work provides real-world in-the-wild measurements that directly correlate 5G radio-layer QoS metrics with objective QoE for YouTube 360degree videos, thereby providing a valuable dataset to support future QoE modeling efforts. III. EMPIRICAL ANALYSIS WITH REAL-WORLD DATA Data Collection Methodology. The dataset was collected using commercial smartphones and the GNetTrack Pro1tool, following the methodology in our previous work [12]. In this study, we focused on 360-degree YouTube videos, recording both channel metrics and QoE logs at one-second granularity. The dataset with related information, such as video selection details, is available on GitHub.2Data was gathered under two scenarios: indoor (static), conducted in the city hub, and mobility, covering more than 100 miles of driving. Mobility traces included variations in speed, ranging from traffic stops to highway speeds of up to 65 mph. Analysis of YouTube QoE and Channel Metrics. Figure 1(a–c) presents 25 indoor traces from Operator X, showing variations in RSRP, RSRQ, and SNR. In comparison, Figure 1(d–f) illustrates 25 mobility traces, including 10 from Operator X and 15 from Operator Y, with a red line separating the operators for clarity. Note that the dataset collected during mobility was more limited than for the indoor scenario, due to practical constraints on trace collection. Figure 2 depicts 30 additional indoor traces from Operator Y, further highlighting variability in channel conditions across operators. Figures 3–5 show the distribution of video bytes downloaded (VBD) across playback resolutions, complementing the 1https://gyokovsolutions.com/g-nettrack/ 2https://github.com/razaulmustafa852/5G360 TABLE I OBJECTIVE QOE RESOLUTIONS (QUALITY SHIFTS)IN %ACROSS STATIC AND MOBILITY PATTERNS FOR OPERATORS, X AND Y Static Mobility Qualities XYXY hd2160 1.73 0.82 1.83 - hd1440 36.24 62.45 76.61 79.56 hd1080 18.29 31.98 15.81 20.42 hd720 10.54 0.20 0.38 0.01 large 32.10 4.39 5.34 - medium 1.05 - - - small 0.02 0.02 - - tiny - 0.10 - - channel metrics in Figures 1 and 2. In all cases, the x-axis represents the experiment number, where each bar corresponds to a single streaming session. Figure 3 corresponds to the indoor traces of Figure 1(a–c) for Operator X, Figure 4 to the mobility traces of Figure 1(d–f) for Operators X and Y, and Figure 5 to the indoor traces of Operator Y in Figure 2. In the indoor case (Figure 3), the quality distribution varied across experiments, with most sessions dominated by hd1440 and hd1080, and some including portions of hd720 or lower resolutions. These fluctuations reflect the impact of weaker indoor channel conditions observed in Figure 1(a–c). For example, in the first indoor experiment, the session was split between hd1440 and hd1080, while approximately 50% of the corresponding channel metrics fell within the ranges of RSRP (–103 to –107 dBm), RSRQ (–12 to –13 dB), and SNR (5 to 7 dB). Note that white gaps observed in some experiments (e.g., experiments 3 and 4 in Figure 3) correspond to slight variations in VBD across consecutive chunks, caused by differences in scene complexity (e.g., static versus highmotion content) or buffer dynamics. In contrast, mobility sessions (Figure 4) were dominated by hd1440 and hd1080, with negligible representation of lower resolutions, consistent with the stronger signal characteristics in Figure 1(d–f). Indoor sessions from Operator Y (Figure 5) demonstrated relatively more stable performance than Operator X, sustaining hd1440 playback for longer periods and occasionally reaching hd2160, in line with the channel conditions shown in Figure 2. Table I quantifies the distribution of playback resolutions across static and mobility scenarios. Higher qualities (hd1440 and hd1080) consistently required more video bytes downloaded (VBD), as expected due to their higher bitrate demands. Notably, mobility scenarios yielded superior QoE compared to indoor sessions, with more than 94% of playback occurring at HD resolutions (hd1080 and above) for both Operators X and Y. This is in contrast to traditional 2D streaming, where mobility often leads to performance degradation [13]. The dataset, available on GitHub3, is well-balanced and captures real-world network conditions, providing a valuable resource for the research community to study the relationship between QoS and QoE and to improve network performance. 3In the shared dataset, operators are anonymized and denoted as X, Y, and Z. Operator Z, while included in the measurement campaign, exhibited irregular behavior and is excluded from the main analysis but remains available in the repository. 1 2 3 4 5 6 7 8 9 101112 13 14 1516171819 20 21 22232425 120 115 110 105 100 95 RSRP (a) Indoor: RSRP 1 2 3 4 5 6 7 8 9 101112 13 14 1516171819 20 21 22232425 18 16 14 12 10 8 RSRQ (b) Indoor: RSRQ 1 2 3 4 5 6 7 8 9 101112 13 14 1516171819 20 21 22232425 10 5 0 5 10 15 SNR (c) Indoor: SNR 1 2 3 4 5 6 7 8 9 101112 13 14 1516171819 20 21 22232425 120 110 100 90 80 70 60 RSRP (d) Mobility: RSRP 1 2 3 4 5 6 7 8 9 101112 13 14 1516171819 20 21 22232425 20 18 16 14 12 10 8 RSRQ (e) Mobility: RSRQ 1 2 3 4 5 6 7 8 9 101112 13 14 1516171819 20 21 22232425 0 10 20 30 40 SNR (f) Mobility: SNR Fig. 1. Channel Metrics – RSRP, RSRQ, SNR for Indoor and Mobility Traces – (Operator X and Y). 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930 120 115 110 105 100 95 RSRP (a) Indoor: RSRP 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930 18 16 14 12 10 RSRQ (b) Indoor: RSRQ 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930 10 5 0 5 10 15 SNR (c) Indoor: SNR Fig. 2. Channel Metrics – RSRP, RSRQ, SNR for Indoor Traces – (Operator Y). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 0.0 0.2 0.4 0.6 0.8 1.0 Video Bytes Downloaded hd2160 hd1440 hd1080 hd720 large medium small Fig. 3. YouTube Qualities During Streaming Sessions for Indoor Traces – (Operator X). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 0.0 0.2 0.4 0.6 0.8 1.0 Video Bytes Downloaded hd2160 hd1440 hd1080 hd720 large medium small Fig. 4. YouTube Qualities During Streaming Sessions for Mobility Traces – (Operator X and Y). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 0.0 0.2 0.4 0.6 0.8 1.0 Video Bytes Downloaded hd2160 hd1440 hd1080 hd720 large medium small tiny Fig. 5. YouTube Qualities During Streaming Sessions for Indoor Traces – (Operator Y). IV. CONCLUSION In this work, we presented a dataset of 360-degree video traffic collected in the wild. It offers one-second granularity of both channel-level QoS metrics and YouTube objective QoE indicators. Measurements were conducted across three major U.S. mobile operators under both static (indoor) and mobility use cases, capturing diverse network conditions. Our current study has certain limitations. The dataset is restricted to three operators and specific locations; broader geographic and operator coverage would improve generalization. Moreover, as the data focuses on Non-Standalone (NSA) 5G networks, future work should also consider Standalone (SA) deployments. As ongoing work, we are extending this dataset in several directions. First, we plan to integrate the collected traces into emulation environments for repeatable and reproducible testing. Second, we are developing a Generative Adversarial Network (GAN) to synthesize additional data, addressing the limitations of field-collected traces. While we have obtained preliminary results from this effort, space constraints prevent us from including them in this paper. Furthermore, we aim to investigate correlations between channel performance and additional objective QoE metrics and apply ML models for predictive QoE estimation. These steps will support operators and researchers in enhancing service-level agreements and optimizing immersive streaming performance. REFERENCES [1] Sandvine, “Global Internet Phenomena Report,” 2024. [Online; accessed 11-May-2025]. [2] V. D. Hooft et al., “A Tutorial on Immersive Video Delivery: From Omnidirectional Video to Holography,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1336–1375, 2023. [3] R. I. T. D. C. Filho et al., “Dissecting the Performance of VR Video Streaming through the VR-EXP Experimentation Platform,” ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), vol. 15, no. 4, pp. 1–23, 2019. [4] M. T. Islam et al., “Predicting XR services QoE with ML: Insights from in-band encrypted QoS features in 360-VR,” in 2023 IEEE 9th International Conference on Network Softwarization (NetSoft), pp. 80– 88, IEEE, 2023. [5] M. Z. Shafiq et al., “Understanding the Impact of Network Dynamics on Mobile Video User Engagement,” ACM SIGMETRICS Performance Evaluation Review, vol. 42, no. 1, pp. 367–379, 2014. [6] C. Kattadige et al., “360NorVic: 360-Degree Video Classification from Mobile Encrypted Video Traffic,” in Proceedings of the 31st ACM Workshop on Network and Operating Systems Support for Digital Audio and Video, pp. 58–65, 2021. [7] J. Ruan and D. Xie, “A Survey on QoE-Oriented VR Video Streaming: Some Research Issues and Challenges,” Electronics, vol. 10, no. 17, p. 2155, 2021. [8] O. Pe˜ naherrera et al., “ML-based Communications Management for XR Services,” Authorea Preprints, 2023. [9] O. S. Pe˜ naherrera-Pulla et al., “ML-Powered KQI Estimation for XR Services: A Case Study on 360-Video,” IEEE Open Journal of the Communications Society, 2024. [10] A. Telili et al., “360-Degree Video Super Resolution and Quality Enhancement Challenge: Methods and Results,” arXiv preprint arXiv:2411.06738, 2024. [11] G. Kougioumtzidis et al., “Deep Reinforcement Learning-Based Resource Allocation for QoE Enhancement in Wireless VR Communications,” IEEE Access, 2025. [12] R. U. Mustafa et al., “YouTube goes 5G: QoE Benchmarking and MLbased Stall Prediction,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC), pp. 01–06, IEEE, 2024. [13] R. U. Mustafa et al., “Investigating the Impact of Channel Metrics in 5G NSA and SA on Video Streaming QoE,” in NOMS 2025-2025 IEEE Network Operations and Management Symposium, pp. 1–7, IEEE, 2025.