Related Experiment Videos
UVMulti: A Large-scale Benchmark Dataset for Underwater Video-level Multi-task Learning
Yuxin Li1, Yuxiang Zhang1, Zhibin Yu2,3
1College of Electronic Engineering, Ocean University of China, Qingdao, 266520, Shandong, China.
Scientific Data
|May 11, 2026
Summary
Researchers introduce UVMulti, the first large-scale underwater dataset for multi-task learning. This benchmark aids developing advanced underwater vision systems by providing diverse annotations and a novel framework.
Area of Science:
- Computer Vision
- Robotics
- Marine Technology
Background:
- Multi-task learning shows promise for underwater vision tasks.
- Lack of large-scale, annotated datasets hinders underwater multi-task learning research.
Purpose of the Study:
- Introduce UVMulti, the first large-scale, high-resolution underwater video benchmark for multi-task learning.
- Provide a comprehensive dataset with pixel-level segmentation, image enhancement, and depth annotations.
Main Methods:
- Developed UVMT-Net, a multi-task learning framework integrating various paradigms.
- Implemented an adaptive task weight adjustment (AWA) strategy to optimize performance.
Main Results:
- The UVMulti dataset contains 100 video sequences (87,352 frames).
- UVMT-Net framework effectively utilizes limited annotated data for improved multi-task performance.
- AWA strategy enhances main task performance while preserving auxiliary task performance.
Conclusions:
- UVMulti serves as a robust benchmark for advancing underwater multi-task learning.
- The proposed methods offer a viable approach for real-world underwater vision applications.
