YOLO Nano Underwater: A Fast and Compact Object Detector for Embedded Device
Lin Wang, Xiufen Ye, Huiming Xing, Zhengyang Wang, Peng Li · Global Oceans 2020: Singapore – U.S. Gulf Coast · 2020
Recent researches on Underwater object detection have progressed with the development of deep learning methods. A large portion of ROVs and AUVs are working on constrained environments with limited power supply and computing capability. In this paper, we propose a fast and compact object detector for several marine products, such as scallop, starfish, echinus and holothurian. The novel proposed model named YOLO Nano Underwater is modified based on YOLO Nano architecture to reduce the inference time. And instance normalization is introduced to replace the batch normalization in some early layers for a precision boost. The model achieves a competitive performance on edge computing devices, like NVIDIA Jetson Nano. Comparing to YOLO v3, our model can get competitive performance in precision, but with only 5% computation. Meanwhile, we re-annotated the training part of the URPC 2019 dataset and the refined annotations can be public available at https://github.com/wangsssky/Refined-training-set-of-URPC2019/.