Hyunseung and Sheetal demonstrated ApproxBit at the NSF CPS PI meeting in Vanderbilt on March 13-14, 2025. This is a part of our NSF Center, “CHORUS: Resilient Distributed CPS through Rational and Dynamic Decision-Making Among Multiple Stakeholders.“
In today’s digital video era, recognizing human actions efficiently from videos is crucial for applications such as surveillance, autonomous driving, and augmented reality. However, deploying video recognition models on resource-constrained edge devices and in low-bandwidth environments remains a significant challenge. Traditional approaches rely on offloading computations from edge devices to servers, but these methods struggle to handle the high data rates associated with video analytics.
To address this issue, ApproxBit introduces an optimized framework for shared edge-to-cloud processing, specifically designed for video action recognition and video question answering (QA). The framework integrates an encoder within the video recognition model and utilizes learned binary codes to efficiently compress and offload data. By adaptively selecting the offloading point based on network bandwidth, ApproxBit significantly improves processing efficiency and reduces latency.
One of the key innovations of ApproxBit is its adaptive data compression, which reduces the original feature map size by up to 1536x. This substantial reduction makes ApproxBit an ideal solution for video action recognition on edge devices, particularly in constrained network conditions. The system has been evaluated on benchmark datasets such as Something-Something-v2 and Kinetics, demonstrating superior performance in both latency and accuracy compared to several baselines: edge-only processing, server-only processing, full feature map offloading, lossy compression methods, H.264-encoded video offloading, JPEG-compressed feature maps, as well as DeepCOD (SenSys ’20) and LimitNet (MobiSys ’24).
ApproxBit’s capabilities showcase its potential to enable efficient and scalable video analytics in real-world settings, bridging the gap between edge and cloud computing for resource-intensive tasks.