Salient Object Detection in the Deep Learning Era: An In-Depth Survey.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Salient Object Detection (SOD) is a critical computer vision task.
• Deep learning (deep SOD) has driven recent advancements in SOD.

Purpose of the Study:

• To provide a comprehensive survey of deep SOD algorithms.
• To analyze existing SOD datasets, evaluation metrics, and model performance.
• To investigate SOD model robustness and identify future research directions.

Main Methods:

• Reviewing deep SOD algorithms based on network architecture, supervision, learning paradigm, and detection level.
• Summarizing and analyzing SOD datasets and evaluation metrics.
• Benchmarking representative SOD models and analyzing performance under attribute settings.
• Constructing a novel SOD dataset with attribute annotations.
• Analyzing model robustness against perturbations and adversarial attacks.
• Examining SOD dataset generalization and difficulty.

Main Results:

• A detailed taxonomy of deep SOD algorithms is presented.
• Existing SOD datasets and evaluation metrics are summarized and analyzed.
• Performance benchmarks of representative SOD models are provided.
• A novel SOD dataset with rich attribute annotations is introduced.
• The robustness of SOD models to various perturbations is analyzed for the first time.
• Insights into dataset generalization and difficulty are offered.

Conclusions:

• The survey provides a thorough understanding of the deep SOD field.
• The newly constructed dataset and robustness analyses offer valuable resources for future research.
• Open issues and future research directions in salient object detection are outlined.