Uncertainty-Aware Active Domain Adaptive Salient Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning significantly advanced salient object detection (SOD).
• Current deep learning SOD methods require extensive pixel-wise annotations, posing a data annotation burden.
• Existing weakly-supervised and unsupervised SOD methods show a performance gap compared to fully supervised approaches.

Purpose of the Study:

• To propose a novel, cost-efficient salient object detection (SOD) framework.
• To adapt SOD models from synthetic to real-world data using limited, actively selected annotations.
• To bridge the performance gap between weakly-supervised/unsupervised SOD and fully supervised methods.

Main Methods:

• Constructed a synthetic SOD dataset by copying and pasting foreground objects onto backgrounds.
• Developed an uncertainty-aware active domain adaptation algorithm for real-world data labeling.
• Utilized prediction variances against data augmentations to compute superpixel-level uncertainty.
• Generated pseudo-labels for low-uncertainty superpixels and manually labeled high-uncertainty ones.

Main Results:

• The proposed framework effectively transfers models from synthetic to real-world SOD datasets.
• The active domain adaptation strategy generates high-quality labels with minimal annotation cost.
• Experimental results on six benchmark SOD datasets show superior performance over existing weakly-supervised and unsupervised methods.
• Achieved performance comparable to fully supervised SOD methods.

Conclusions:

• The proposed cost-efficient framework successfully addresses the annotation burden in deep learning-based SOD.
• Active domain adaptation with uncertainty estimation is a viable strategy for improving SOD performance with limited annotations.
• The method offers a practical solution for achieving high-performance SOD in real-world scenarios without extensive manual labeling.