The Internet is moving to edge computing architectures to accommodate latency-sensitive DNN workloads in the developing Internet of Things and mobile computing application domains. Unfortunately, large-scale, high-precision DNN models cannot work at the edge due to the lack of computing power, unlike cloud environments. Efforts to date have therefore focused on moving some computing to the cloud to get around this limitation. However, this will result in longer delays.
New research from Microsoft proposes REACT, a unique architecture that uses the edge and cloud together to perform redundant computation. Asynchronously received cloud inputs are fused into the computational stream at the edge to improve detection quality without compromising latency. This allows you to take advantage of the accuracy of the cloud without sacrificing low edge latency.
The team uses two approaches to solve the problem of lack of edge computing power and accuracy loss due to edge models.
- First, due to the spatio-temporal correlation between consecutive video frames, edge object identification needs to be called only once every few frames. Edge detection is done every 5 frames. It uses a fairly lightweight operation of object tracking to bridge the gap between the two framesets.
- Then, only certain frames are asynchronously sent to the cloud for better inference accuracy. Depending on network latency and availability of cloud resources, the edge device may not get cloud detection for several subsequent frames.
- The latest, previously unreported cloud detections are then combined with the current imagery. To “fast-forward” to the current time, use the cloud detection generated in the old frame and feed it to her second instance of the object tracker. Newly identified items can be merged into the current frame as long as there is no dramatic change in the scene.
The team applied this method to a dataset of dashcam video. Their experiments used state-of-the-art computer vision techniques to obtain local and remote item detection. In addition, we evaluate the quality of object detection using a widely used statistic in the field of computer vision known as mAP@0.5 (average accuracy at 0.5 IoU). They also looked at two datasets by him to determine how effective REACT was.
- VisDrone as a drone-based surveillance system
- The D2City system is a driving support system based on a drive recorder.
Results from their tests show that REACT can provide up to 50% better results than baseline methods. It also demonstrates that edge and cloud models can complement each other, and that the proposed edge-cloud convergence approach can improve overall performance.
In addition to the light object tracking which is done in intermediate frames, the object detector is run only once every few frames. By overlapping detection between edge and cloud, developers have more freedom to choose how often their applications run on each platform while maintaining the same level of detection accuracy.
The researchers also emphasize that when multiple edge devices use the same cloud-hosted model, the cost of using cloud resources may be spread across more people. In particular, assuming your application can tolerate a central latency of up to 500ms, the V100 GPU can support more than 60 concurrent devices simultaneously.
Although this work has primarily described its application to object detection, the team believes it can be applied to other situations, such as human pose estimation, instance, and semantic segmentation applications to achieve the “best of both worlds.” I’m here.
Please check paper and article.don’t forget to join 21,000+ ML SubReddit, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email me. Asif@marktechpost.com
🚀 Check out 100’s of AI Tools at the AI Tools Club
Tanushree Shenwai is a consulting intern at MarktechPost. She is currently pursuing her bachelor’s degree at the Indian Institute of Technology (IIT), Bhubaneswar. She is a data her science enthusiast and has a keen interest in the range of applications of artificial intelligence in various fields. She is passionate about exploring new advances in technology and its practical applications.
