Mini-YOLOX: A Lightweight Network for Real-Time Embedded Applications.

In this paper, we present an efficient real-time object detector designed for embedded system applications. Although increased memory and power consumption pose significant challenges, our network adopts techniques to overcome them. Depthwise and pointwise separable convolutions are used across the backbone, neck, and head of the network to reduce the number of parameters. To compensate for accuracy loss, we employ several techniques including starting with a larger baseline, using depthwise separable convolution mixed with blueprint separable convolution in the backbone, replacing the SiLU activation function with the Mish activation function, and switching from Adam to SGD optimization algorithms. These techniques help to achieve a state-of-the-art object detection network called Mini-YOLOX, which optimizes YOLOX-S from the YOLOX series and reaches only ∽57% of its parameter size and ∽52% of its floating-point operations with better accuracy. Mini-YOLOX achieves impressive results with only 5.08M parameters and 13.95 GFLOPs, resulting in a 40.9 AP score in just 5.71 ms.