Revolutionizing Edge Computing: REACT Unveils Cloud-Assisted Approach for Reduced Latency & Advanced DNN Workloads

As the Internet of Things (IoT) and mobile computing applications continue to expand, managing latency-sensitive deep neural network (DNN) workloads has become increasingly critical. Edge computing devices were introduced to address these challenges, but their limitations often resulted in shifted tasks to the cloud, which ironically increased latency.

Revolutionizing this landscape is REACT, an architecture that seamlessly merges edge and cloud computing for redundant calculations. REACT aims to enhance accuracy without sacrificing latency, using cloud inputs to asynchronously integrate into the computation stream at the edge.

The Two-Pronged Approach: Frame Skipping and Asynchronous Fusion

REACT’s innovative approach relies on two primary techniques: frame skipping, and asynchronous fusion of cloud detections.

1. Frame Skipping:

Leveraging spatial-temporal correlations, edge object detection is called only once every few frames, significantly reducing latency. To maintain accuracy, a lightweight object tracking operation occurs every fifth frame, bridging the gap between detection calls.

2. Asynchronous Fusion of Cloud Detections:

To further refine accuracy, certain frames are transmitted to the cloud, while cloud detections are received after a few frames’ delay. The system then combines the most recent cloud detections with the current image using object tracking, resulting in a robust and accurate detection process.

Experimental Setup and Datasets:

Realizing the potential of REACT, researchers conducted experiments using dashcam video datasets along with top-of-the-line computer vision techniques for obtaining both local and remote object detections. The measure of mAP@0.5 was employed to evaluate detection quality, with the VisDrone and D2City datasets providing a comprehensive playground for testing REACT’s effectiveness.

Promising Results and a Bright Future for REACT:

The experiments’ results were undeniably impressive, showcasing that REACT can offer up to 50% better performance than baseline methods. These findings indicate that edge and cloud models can produce even greater results when working together in a complementary manner.

The proposed edge-cloud fusion approach not only paves the way for improved overall system performance, but also ushers in a new era of possibilities for edge computing solutions. As industries continue to pursue seamless connections in IoT and mobile computing applications, REACT serves as a beacon of hope for managing latency-sensitive DNN workloads without compromising precision and efficiency.