UVTD: A Large-Scale Multilabel Data Set for Underwater Vision Tasks

With the exploration of ocean resources, underwater vision tasks (UVTs) have attracted increasing interest in recent years. However, the advancement of UVTs is hampered by several challenges, prominently the difficulty in acquiring large-scale data sets with accurate annotations, and the absence of a unified multi-label data set for underwater multitask learning (UMTL). Motivated by the critical need for a large-scale, multilabel data set tailored for UVTs, we present a large-scale multilabel data set for underwater vision tasks (UVTD), which offers solutions to several bottlenecks in UVT research: first, supporting diverse applications. With annotations for two underwater low-level vision tasks (i.e., underwater image enhancement and underwater image quality assessment) and three high-level vision tasks (i.e., underwater semantic segmentation, underwater object detection, and underwater salient object detection), UVTD supports a wide range of underwater applications. Second, enhancing model generalization. UVTD comprises 5380 real underwater images, covering diverse scenes and varied degradation characteristics, improving the robustness of vision algorithms. Finally, facilitating multitask learning. Our multilabel data set enables researchers to explore the correlations between tasks and develop robust UMTL algorithms. Based on UVTD, we propose two UMTL networks tailored to the low-level and high-level tasks separately, serving as benchmarks for future research in the UMTL field. Extensive experiments demonstrate UVTD's superiority across multiple UVTs, and the proposed UMTL networks exhibit competitive performance on these tasks, implying the significant implications of UVTD for future research.