Salient Objects in Clutter.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing salient object detection (SOD) datasets exhibit a design bias, assuming clear, uncluttered salient objects.
• This bias leads to saturated performance on benchmarks but poor real-world applicability of SOD models.

Purpose of the Study:

• To address the design bias in current SOD datasets.
• To propose a new high-quality dataset and update the saliency benchmark for more realistic evaluations.
• To investigate dataset enhancement strategies for improving SOD model performance.

Main Methods:

• Introduction of the Salient Objects in Clutter (SOC) dataset, featuring images with salient and non-salient objects in cluttered scenes.
• Inclusion of object category annotations and challenge-related attributes in the SOC dataset.
• Investigation of dataset enhancement strategies: label smoothing, random image augmentation, and self-supervised learning.

Main Results:

• The proposed SOC dataset and enhancement strategies significantly improve SOD model performance.
• Label smoothing, augmentation, and self-supervised learning demonstrate effectiveness in addressing SOD challenges.
• The study provides a comprehensive benchmark for SOD research.

Conclusions:

• Improving datasets is crucial for advancing salient object detection beyond current limitations.
• The SOC dataset and proposed methods offer a more robust evaluation and training paradigm for SOD models.
• The findings pave the way for more effective SOD applications in real-world environments.