Dynamic Bandwidth Allocation for Hybrid Event-RGB Transmission

Pujing Yang, Guangyi Zhang, Yunlong Cai, Lei Yu, Guanding YuarXiv 2025 · DOI · arXiv

Introduction

Hybrid camera systems, which integrate standard RGB sensors with bio-inspired event cameras, represent a significant advancement for high-speed computer vision. While RGB cameras provide essential color and texture information, they are prone to motion blur. Conversely, event cameras capture pixel-level intensity changes with microsecond latency but lack static detail. Despite their complementary nature, these systems face a critical data bottleneck. Event cameras generate high volumes of data during rapid motion, and because both sensors observe the same scene, they often capture redundant information. Traditional transmission methods fail to exploit this redundancy, leading to inefficient bandwidth usage.

The Joint Event-Image Transmission (JEIT) Framework

To address these challenges, the researchers propose the Joint Event-Image Transmission (JEIT) framework. Moving beyond traditional "compress-then-transmit" paradigms, JEIT employs Deep Joint Source-Channel Coding (JSCC) to optimize data flow from hybrid sensors.

The JEIT framework disentangles shared and specific information from hybrid sensors to optimize bandwidth allocation.

The system is designed to dynamically shift bandwidth to the sensor most relevant to the current scene. In high-motion scenarios, bandwidth is prioritized for the event stream to capture dynamics; in static scenes, the RGB stream is prioritized for detail. This is achieved through specialized fusion modules: the U-shape Feature Fuse Module (UFFM) for multi-scale feature integration and the Cross Attention Fuse Module (CAFM) for adaptive weighting during reconstruction.

Information Disentanglement

A core innovation of JEIT is its ability to disentangle shared information from domain-specific details. Using the Information Bottleneck (IB) principle and Bayesian modeling, the system separates input data into three latent categories: image-specific, event-specific, and shared information.

Bayesian graphical models illustrating the information compression and preservation stages of the disentanglement process.

As shown in the graphical models above, this disentanglement ensures that redundant features—such as object shapes captured by both sensors—are not transmitted twice. By estimating the entropy of these latent representations, the system achieves "semantic" bandwidth allocation, transmitting only the most meaningful bits required for high-quality reconstruction and deblurring. Unlike pipelines that reconstruct and then deblur, JEIT integrates deblurring into the decoding process, accelerating the pipeline for real-time edge-to-server applications.

Experimental Evaluation

The JEIT framework was evaluated on the GoPro and REBlur datasets. The GoPro dataset provided synthetic event data via the ESIM simulator, while REBlur offered real-world synchronized event-RGB pairs captured from hardware.

Key Results

  • Bandwidth Efficiency: JEIT achieved approximately 17% bandwidth savings compared to separate transmission benchmarks like EFNet combined with DeepJSCC.
  • Superior Deblurring: The system maintained high PSNR even at low bandwidth ratios, effectively preserving sharp edges and fine textures (such as text on moving vehicles) where other methods produced artifacts.
  • Robustness: JEIT demonstrated resilience to wireless noise across various Signal-to-Noise Ratios (SNR) in Additive White Gaussian Noise (AWGN) channels.

Conclusion

JEIT demonstrates that treating sensor correlation as a resource rather than a nuisance can significantly improve transmission efficiency and task performance in hybrid vision systems. While the current model shows significant gains, future work will focus on reducing the computational footprint of the fusion modules and testing robustness against more complex real-world channel models involving fading or multi-path interference.

Summary updated

BibTeX
@misc{yang2025dynamic,
  author = {Pujing Yang and Guangyi Zhang and Yunlong Cai and Lei Yu and Guanding Yu},
  title = {{Dynamic Bandwidth Allocation for Hybrid Event-RGB Transmission}},
  year = {2025},
  eprint = {2506.20222},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.20222},
}
Figure