Edge-guided oceanic scene element detection

Abstract

Oceanic environmental monitoring is critical to environmental protection. As a core technique, Oceanic Scene Element Detection (OSED) plays an important role. Existing oceanic object detection approaches are usually focused on a single category. Therefore, a multi-category OSED data set is demanded. Considering oceanic scene elements normally present large-scale complicated structures, the edge cue is particular useful for representation of these elements. However, none of existing object detection methods take this cue into account. To address the two problems, we first collect and annotate three OSED data sets, which comprise a total of 10,040 images and 60 categories. Then we propose a generic Multi-scale Edge-Guided Module (MSEGM), which can be inserted into an object detection network, for guiding the backbone toward learning edge characteristics. An Edge-Guided Oceanic Scene Element Detection (EG-OSED) framework is built on top of this module and a base object detector, which can be end-to-end trained using a multi-task learning scheme. A series of experiments are conducted on the three OSED data sets. The results demonstrate that the EG-OSED framework normally outperforms its base object detector which does not utilize edges. We believe that these promising results should be due to the importance of edges to representation of oceanic scene elements.