Underwater object detection in noisy imbalanced datasets

Long Chen, Tengyue Li, Andy Zhou, Shengke Wang, Junyu Dong, Huiyu Zhou

Abstract: Highlights•We provide theoretical analysis and empirical analysis on what factors result in the imbalanced detection problem. As we know, it is the first work to conduct the theoretical analysis.•We propose a factor-agnostic gradient re-weighting (FAGR) algorithm to address the imbalance problem. FAGR produces the precision balanced gradients for all the classes and rebalances the precision distributions.•Extensive experiments show that our proposed framework, by re-balancing precision distributions, performs much better than those of re-balancing data distributions.