TripleNet: Exploiting Complementary Features and Pseudo-Labels for Semi-Supervised Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised salient object detection faces challenges due to limited labeled data and adapting conventional strategies.
• Existing methods struggle with the inherent limitations of output categories in salient object detection tasks.

Purpose of the Study:

• To propose a novel multi-branch architecture, TripleNet, for semi-supervised salient object detection.
• To address the limitations of conventional semi-supervised strategies by extracting complementary features from limited labeled data.

Main Methods:

• Introduced TripleNet, a three-branch network for contour, content, and holistic saliency prediction.
• Decomposed limited ground truths to derive supervision signals for contour and content branches.
• Developed coupled and enhanced pseudo-labeling mechanisms using complementary features and reliable regions.
• Incorporated a partial binary cross-entropy loss with adaptive thresholding for saliency branch learning.

Main Results:

• Achieved state-of-the-art performance in salient object detection using only 329 labeled training images.
• Demonstrated the effectiveness of the multi-branch architecture in leveraging complementary features.
• Validated the proposed pseudo-labeling strategies and loss function for improved accuracy.

Conclusions:

• The proposed TripleNet architecture and enhanced pseudo-labeling mechanism significantly advance semi-supervised salient object detection.
• Effective utilization of limited labeled data is achieved through complementary feature extraction and novel labeling strategies.
• The method offers a promising solution for accurate salient object detection with minimal labeled data.