Underwater Object Detection in Noisy Imbalanced Datasets

Class imbalance occurs in the datasets with a disproportionate ratio of observations. The class imbalance problem drives the detection and classification systems to be more biased towards the over-represented classes while the under-represented classes may not receive sufficient learning. Previous works often deploy distribution based re-balancing approaches to address this problem. However, these established techniques do not work properly for underwater object detection where label noise commonly exists. In our experiments, we observe that the imbalanced detection problem may be caused by imbalance data distributions or label noise. To deal with these challenges, we first propose a noise removal (NR) algorithm to remove label noise in the datasets, and then propose a factor-agnostic gradient re-weighting algorithm (FAGR) to address the imbalanced detection problem. FAGR provides a rebalanced gradient to each class, which encourages the detection network to treat all the classes equally whilst minimising the detection discrepancy. Our proposed NR+FAGR framework achieves state-of-the-art (SOAT) performance on three underwater object datasets due to its high capacity in handling the class imbalance and noise issues. The source code will be made available at: https://github.com/IanDragon.