Source: Deep Learning on Medium
YOLO Nano: A Highly Compact You Only Look Once Network for Object Detection
One of the most cutting-edge computer vision applications, object detection is increasingly drawing a lot of interest from the research community. Deep learning, especially the use of deep convolutional neural networks being the force behind its recent advances. Object detection demands model accuracy which involves a lot of complex training. Models such as R-CNN and other improved versions brought efficiency in the field. But then YOLO came and beat R-CNN and its all variants.
Conventional object detection models have demonstrated state-of-the-art performance. But it is very challenging, or otherwise impossible to deploy on edge and mobile devices due to computational and memory constraints. Faster R-CNN, for instance, have inference speeds at low single-digit frame rates on embedded processors.
It’s important to address the challenge of achieving embedded object detection through the design of highly efficient model architectures that are more well-suited for edge and mobile devices.
YOLO Nano for Object Detection
With a growing interest to design and develop better for object detection especially for mobile technology, researchers have introduced YOLO Nano, a highly compact deep CNN for the task of object detection. To make this happen, they leveraged a human-machine collaborative design strategy that comprises principled network design prototyping, and machine-driven design exploration. YOLO Nano has a model size of ~4.0MB and needs 4.57B operations for inference while still achieving an mAP of ~69.1% on the VOC 2007 dataset.
Potential Uses and Effects
Researchers examined YOLO Nano model size, accuracy, and computational cost on the PASCAL VOC datasets. From the paper, results demonstrate the efficacy of YOLO Nano for embedded scenarios on inference speed and power efficiency.
I think YOLO is a great step towards achieving robust systems where local embedded processing is required such as video surveillance, IoT, unmanned aerial vehicles, autonomous driving, and more.