A Comparative Analysis of the Dataset for Training Underwater Fish Detector based on YOLOv3

抄録

A spectacular diversity of fishes under a crystal clear seawater in Okinawa attracts numerous scuba divers, snorkelers around the world. With the advancement in computer vision and deep learning, object detection is much more reliable than ever and find its application almost in every industry, and also in marine leisure activity. Being able to detect and recognize all underwater objects provides both an educational and amazing experience to divers and snorkelers to explore the underworld. However, it requires a system that could work in real-time with high accuracy. This is a challenge that all deep learning-based object detection algorithm is facing since there is a trade-off between time and accuracy. YOLOv3 is one of the fastest object detection algorithms that can work in real-time. We use this to train and test on our custom dataset. We collected the underwater fish image and built our dataset that contains 3548 images. We provide a comparative analysis of the training and evaluation of three different datasets. With data augmentation, our model can achieve up to 92% of mAP, and we also show what role that negative data impact the performance of the model.