USOD10K: A New Benchmark Dataset for Underwater Salient Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Image Processing

Background:

• Underwater salient object detection (USOD) is crucial for underwater tasks but lacks large-scale, well-annotated datasets.
• Existing datasets are limited in diversity, complexity, and scalability, hindering USOD research progress.

Purpose of the Study:

• Introduce USOD10K, a novel, large-scale dataset for underwater salient object detection.
• Propose and evaluate a new baseline model, TC-USOD, for improved USOD performance.
• Benchmark existing methods and provide a foundation for future USOD research.

Main Methods:

• Developed USOD10K dataset with 10,255 images, 70 object categories, 12 scenes, and pixel-wise annotations including depth maps.
• Proposed TC-USOD, a hybrid encoder-decoder model using transformers and convolutions.
• Benchmarked 35 state-of-the-art SOD/USOD methods on existing and the new USOD10K dataset.

Main Results:

• USOD10K is the first large-scale, diverse, and scalable dataset for USOD.
• The proposed TC-USOD model achieved superior performance across all tested datasets.
• Comprehensive benchmarking provides insights into current SOD/USOD method capabilities.

Conclusions:

• The USOD10K dataset and TC-USOD model represent a significant advancement in underwater salient object detection.
• This work will accelerate research in underwater visual tasks and visually-guided underwater robots.
• Publicly available dataset, code, and results aim to foster further development in the USOD field.