Salient object detection: A survey

Analysis

Why This Paper Matters

Salient object detection (SOD) is a fundamental problem in computer vision with wide-ranging applications, from image compression to content-aware editing and visual tracking. This survey, published in 2019, arrives at a critical juncture when deep learning has revolutionized the field, yet a consolidated understanding of the rapid progress was missing. By reviewing 228 publications, the authors provide a much-needed structured overview, helping researchers navigate the crowded landscape of models, datasets, and metrics.

The paper's significance extends beyond mere cataloging. It explicitly situates SOD among related tasks like generic scene segmentation, object proposal generation, and saliency for fixation prediction, clarifying the subtle distinctions and connections. This contextualization is invaluable for interdisciplinary researchers and for avoiding conceptual confusion that often arises when these terms are used interchangeably. The survey also identifies persistent issues such as dataset bias and the inadequacy of existing evaluation metrics, which are crucial for the field's healthy development.

Technical Contributions

The survey's main technical contributions are organizational and analytical:

• Comprehensive taxonomy: It categorizes SOD methods into historical roots (e.g., cognitive science, psychophysics) and modern trends, including deep learning-based approaches.
• Task clarification: It delineates SOD from related tasks, highlighting differences in goals and evaluation protocols.
• Technique review: It systematically covers core techniques, from handcrafted features to end-to-end deep networks, and discusses modeling trends like multi-scale processing and attention mechanisms.
• Dataset and metric analysis: It compiles a list of popular datasets and evaluation metrics, discussing their strengths and weaknesses.
• Open problem identification: It highlights issues like dataset bias and metric limitations, proposing future research directions.

Results

As a survey, the paper does not present new experimental results. Instead, it synthesizes findings from the literature, noting that deep learning-based methods have significantly outperformed traditional approaches on standard benchmarks. It also points out that performance varies across datasets due to bias, and that common metrics like F-measure and MAE have limitations. The survey's value lies in its meta-analysis, offering a bird's-eye view of the field's progress and remaining challenges.

Significance

The impact of this survey is substantial. It has become a widely cited reference (839 citations) for researchers entering the field and for practitioners selecting appropriate methods. By clarifying terminology and consolidating knowledge, it helps reduce redundancy and fosters more targeted research. Its discussion of open problems has likely influenced subsequent work on more robust evaluation metrics and unbiased datasets. Moreover, by linking SOD to broader vision tasks, it encourages cross-pollination of ideas, potentially benefiting areas like object detection and segmentation. For the AI community, this survey exemplifies the importance of periodic synthesis to guide a rapidly evolving field.