UVMulti: A Large-scale Benchmark Dataset for Underwater Video-level Multi-task Learning
Researchers from the Ocean University of China have introduced UVMulti, the first large-scale, high-resolution benchmark dataset designed specifically for underwater video-level multi-task learning. Published in Scientific Data, this initiative addresses the critical shortage of comprehensive datasets with multi-task ground truth annotations that has previously hindered advancements in underwater computer vision. The UVMulti dataset comprises 100 video sequences totaling 87,352 frames, featuring detailed pixel-level segmentation, image enhancement, and depth ground truth annotations. To leverage this data effectively, the team developed UVMT-Net, a multi-task learning framework that integrates various learning paradigms to optimize performance using limited annotated data. Additionally, an adaptive task weight adjustment strategy was implemented to balance main and auxiliary task performance. Extensive experiments confirm UVMulti as a robust and versatile benchmark, significantly promoting research and development in underwater vision tasks. This work is supported by the National Natural Science Foundation of China and the Hainan Province Science and Technology Special Fund, marking a significant contribution to physical oceanography and computer science fields.
Wire timeline
UVMulti: A Large-scale Benchmark Dataset for Underwater Video-level Multi-task Learning
Researchers from the Ocean University of China have introduced UVMulti, the first large-scale, high-resolution benchmark dataset designed specifically for underwater video-level multi-task learning. Published in Scientific Data, this initiative addresses the critical shortage of comprehensive datasets with multi-task ground truth annotations that has previously hindered advancements in underwater computer vision. The UVMulti dataset comprises 100 video sequences totaling 87,352 frames, featuring detailed pixel-level segmentation, image enhancement, and depth ground truth annotations. To leverage this data effectively, the team developed UVMT-Net, a multi-task learning framework that integrates various learning paradigms to optimize performance using limited annotated data. Additionally, an adaptive task weight adjustment strategy was implemented to balance main and auxiliary task performance. Extensive experiments confirm UVMulti as a robust and versatile benchmark, significantly promoting research and development in underwater vision tasks. This work is supported by the National Natural Science Foundation of China and the Hainan Province Science and Technology Special Fund, marking a significant contribution to physical oceanography and computer science fields.
Scientific Data