Abstract
Mobile video conferencing often struggles with bandwidth limitations and small screen sizes. Region of Interest (ROI) transcoding offers a solution by prioritizing the quality of critical areas, such as a speaker's face. However, standard objective quality assessment models like ITU-T G.1070 lack parameters for modern codecs like H.264 and do not account for ROI-based quality distribution. This paper extends the G.1070 model by estimating twelve necessary parameters for H.264 and introducing a weighted quality estimation formula that balances ROI and background quality. Subjective tests validate the model, showing a high correlation (0.978) with human perception.
Introduction: The ROI Transcoding Challenge
Mobile devices face a persistent conflict between limited network bandwidth and the demand for high-quality video. ROI transcoding addresses this by encoding the "important" parts of a frame at higher quality while aggressively compressing the background.
While effective, this technique complicates objective quality assessment. The industry-standard ITU-T G.1070 model was originally designed for older codecs and uniform frame quality. To adapt it for modern mobile conferencing, new parameters for H.264 must be derived, and the model must be modified to integrate disparate quality levels from different regions of the frame.
Methodology
The research modernized the G.1070 model through a three-step process:
- H.264 Parameter Estimation: Using subjective Mean Opinion Scores (MOS) from tests with 20 participants, the authors used Least Square Approximation (LSA) to determine twelve constants ($v_1$ to $v_{12}$) that define the impact of bitrate, frame rate, and packet loss on H.264 video quality.
- ROI Quality Integration: A new overall quality equation was proposed: $V_q = w_1 V_{qBase} + w_2 V_{qROI}$, where $w_1$ and $w_2$ are weighting factors for the background and ROI, respectively.
- Weighting Factor Calculation: Through further subjective validation, the weights were determined to be 0.56 for the ROI and 0.44 for the background, reflecting a user preference for clarity in the primary subject.
Key Insights: The Frame Rate "Sweet Spot"
A critical finding of this study is the non-linear relationship between bitrate, frame rate, and perceived quality.
As shown in the figure above, at low bitrates (e.g., 128–256 kbps), the optimal frame rate is low (6–7 fps). Increasing the frame rate further degrades quality because the per-frame compression becomes too severe. Conversely, at higher bitrates (768 kbps+), the "sweet spot" shifts to 15 fps or higher, as the bandwidth can support more frames without sacrificing individual frame clarity.
Results and Conclusion
The extended G.1070 model demonstrated exceptional performance, achieving a Pearson correlation of 0.978 between estimated scores and subjective MOS. This high accuracy confirms the model's utility as a "plug-and-play" resource for estimating H.264 video quality in ROI-based systems. By prioritizing the ROI with a weight of 0.56, the system aligns closely with human visual attention, providing a more satisfying user experience under constrained network conditions.
Summary updated