BAUODNET for Class Imbalance Learning in Underwater Object Detection

Abstract

Underwater object detection is of great significance for various applications in underwater scenes. However, class imbalance issue is still an unsolved bottleneck for current underwater object detection algorithms. It leads to large discrepancies in the detection precision among different classes that the dominant classes with more training data achieve higher precision while the minority classes with less training data achieve much lower precision. In this paper, we propose a balanced underwater object detection network (BAUODNET) to address the class imbalance issue by exploiting two techniques, i.e., the style augmentation technique and the example re-weighting technique. Firstly, we propose a class-wise style augmentation (CWSA) algorithm to augment the training data for the minority classes that generates different colors, textures and contrasts for the minority classes whilst preserving geometry. The augmented dataset possesses more balanced data distribution; Secondly, we exploit the the focal loss to re-weight the examples during the training of the deep detector, it down-weights the loss assigned to the well-detected examples from the dominant classes and focuses on learning undetected hard examples from the minority classes. Extensive experiments show the effectiveness of CWSA and focal loss for addressing the class imbalance problem in underwater scenes, BAUODNET obtains 49.5% mAP on URPC2017 and 66.8% mAP on URPC2018, achieving state-of-the-art or comparable performance on URPC2017 and URPC2018.