Object Detection on Edge Devices

The battle of object detection based on deep learning algorithms is moving on the edge devices with various types of devices from NVIDIA with Jetson-family [1], Google Coral Board [2]. Even though, current SOTA object algorithms can perform excellently on desktop computers with heavy GPU capacity there is still a large gap in both terms of speed and accuracy when converting and running those models on edge devices. YOLO-Ret [3] with MobileNetV2x0.75 as a backbone network operate real-time on Jetson Nano at around 48 frames-per-second (fps). The object detection algorithms are based on deep learning methods, both in single-stage as well as two-stage methods, rely on multi-scale features. One of the most popular multi-scale architecture is Feature Pyramid Networks (FPN) [4], which created a top-down path from high-level features towards low-level features in order to enhance the accuracy of detection layers using shallower features.