Multiple Early-Exits Strategy for Distributed Deep Neural Network Inference

Existing work on multi-camera network mainly focused on centralized systems, the role of the cameras being limited to capture images and send them to the cloud. However, this approach is impractical when the network connection is limited or variable. Our methodology distributes DNN inference across multiple edge clients and the cloud using early-exits and bottlenecks for time and accuracy requirements. In limited communication scenarios like LoRa, our solution improves average latency by 80 % compared to traditional cloud solutions.