Analysis of Scale Sensitivity of Ship Detection in an Anchor-Free Deep Learning Framework

Abstract

Ship detection is an important task in sea surveillance. In the past decade, deep learning-based methods have been proposed for ship detection from images and videos. Convolutional features are observed to be very effective in representing ship objects. However, the scales of convolution often lead to different capacities of feature representation. It is unclear how the scale influences the performance of deep learning methods in ship detection. To this end, this paper studies the scale sensitivity of ship detection in an anchor-free deep learning framework. Specifically, we employ the classical CenterNet as the base and analyze the influence of the size, the depth, and the fusion strategy of convolution features on multi-scale ship target detection. Experiments show that, for small targets, the features obtained from the top-down path fusion can improve the detection performance more significantly than that from the bottom-up path fusion; on the contrary, the bottom-up path fusion achieves better detection performance on larger targets.