How Perception‑Aware Congestion Control Boosts Real‑Time Video QoE by Up to 32%

Introduction

Real‑time communication (RTC) services such as video conferencing, cloud gaming, and live streaming rely on congestion control to adapt video bitrate and maintain user Quality of Experience (QoE). Existing congestion‑control schemes either focus solely on network metrics or cannot adapt to heterogeneous network environments.

Observations and Motivation

Observation 1: Bandwidth utilization is a non‑decreasing function of queueing delay. In RTC, variable network bandwidth and fluctuating frame sizes create idle periods; increasing bitrate shortens idle periods but raises queueing delay.

Observation 2: Perceived video quality is a non‑decreasing function of bitrate, but the rate of quality increase varies across content types. VMAF curves for different video genres confirm this trend.

These observations motivate a control strategy that simultaneously considers bandwidth utilization, queueing delay, and perceptual video quality.

PACC Architecture

PACC (Perception‑Aware Congestion Control) operates in decision intervals (200 ms). For each interval it estimates:

Perceived video‑quality trend using a CNN‑based Perceptual Video‑Quality Sensor (PVQS).

User delay‑acceptance trend derived from average packet queueing delay.

A combined trend value directs bitrate adjustment toward higher QoE. The algorithm replaces GCC in the Google WebRTC branch (AlphaRTC) and updates the sending bitrate via RTCP feedback.

Perceptual Video Quality Sensor (PVQS)

Directly measuring quality‑vs‑bitrate curves in RTC is impractical. PVQS predicts the quality‑increase rate using a lightweight DenseNet‑121 model trained on VMAF‑labeled video patches. Two bitrate reference points (20 VMAF/Mbps and 2 VMAF/Mbps) define three quality‑rate classes (low, medium, high). The model is fine‑tuned on a small dataset with subjective blockiness and blur scores, then trained on frames from LIVE‑NFLX‑II and MCML datasets, achieving 93.4 % frame‑level accuracy.

Experimental Setup

Experiments run on the AlphaRTC platform with Mahimahi network emulation, reproducing LTE traces and synthetic traces (random queue lengths 10‑600 packets, tail‑drop). Three baseline algorithms are compared:

GCC – model‑based default WebRTC controller.

HRCC – heuristic controller periodically adjusted by an RL agent.

CLCC – pure RL controller optimizing video quality.

Two QoE metrics are used: a network‑level score (weighted combination of delay, bandwidth utilization, and loss) and an application‑level score (weighted sum of frame delay, VMAF quality, and frame loss).

Results

Across 240 experiments (≈6 hours) on 30‑fps, 720p video, PACC consistently outperforms all baselines. On LTE traces, PACC gains QoE of (12.2, 7.5), (4.5, 6.0), and (3.7, 4.2) over HRCC, GCC, and CLCC respectively; on synthetic traces gains reach (17.1, 12.2), (4.4, 4.5), and (4.8, 3.7). PACC also achieves higher bandwidth utilization with negligible queueing‑delay increase. Detailed CDFs show superior performance in all QoE sub‑scores.

Figure 10 illustrates a 90‑second session where PACC quickly stabilizes bitrate, whereas HRCC lags, GCC overshoots bandwidth, and CLCC exhibits large quality oscillations.

Ablation Study

To isolate PVQS impact, three variants are tested:

Random Sensor (RS) – replaces PVQS with a random classifier.

Precise Sensor (PS) – uses ground‑truth quality‑increase rates (unavailable in practice).

PACC with PVQS (original).

Results on a 2.5‑hour trace show PVQS performance close to the ideal PS and far superior to RS, confirming the sensor’s effectiveness.

Conclusion

PACC demonstrates that incorporating perception‑aware metrics—video‑quality trends and user delay‑acceptance—into congestion control yields significant QoE improvements in real‑time video streaming. The lightweight CNN‑based PVQS enables real‑time deployment, and the trend‑based control logic adapts swiftly to dynamic network conditions.