GSDCN: A Customized Two-Stage Neural Network for Benthonic Organism Detection

Zhaoliang Wan, Lu Zhang, Hai Huang, Xu Yang

Abstract: High-quality detection of the benthonic organism is a crucial step to implement autonomous picking for the underwater robot. But there have been few studies on the underwater organism detection in recent years. Directly fine-tuning the generic object detector on an underwater dataset is limited by the domain shift, and thus cannot achieve a good performance. Then we propose a customized two-stage detector named by GSDCN and featured Guided Anchoring mechanism, Sampling Balanced strategy, and Deformable Convolutional module, which is dedicated to overcoming three challenges, i.e., geometric variations, limited underwater visibility range, and the imbalance of object samples. Extensive experiments conducted on the URPC2018 dataset (Publicly available in http://www.cnurpc.org/index.html .) show that our GSDCN improves the detection mAP of our baseline algorithm by 3.40%, and surpasses the state-of-the-art underwater object detector ROIMix [12] by a large margin to 5.39%.