Segment Concealed Objects With Incomplete Supervision.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Incompletely-Supervised Concealed Object Segmentation (ISCOS) is challenging due to limited annotations and difficulty distinguishing objects from backgrounds.
• Existing methods struggle with the inherent similarities between concealed objects and their environments.

Purpose of the Study:

• To propose the first unified method for ISCOS that addresses both limited supervision and object-background similarity.
• To enhance the performance of segmentation models in scenarios with incomplete annotations.

Main Methods:

• A unified mean-teacher framework, SEE, is introduced, leveraging the Segment Anything Model (SAM) for pseudo-label generation.
• Strategies for pseudo-label generation, storage, and supervision are developed to ensure robust training.
• A hybrid-granularity feature grouping module is designed to improve segmentation coherence by clustering similar features.

Main Results:

• The proposed method achieves state-of-the-art performance across multiple ISCOS tasks.
• Experimental results validate the effectiveness of SEE in segmenting concealed objects.
• SEE demonstrates significant improvements in segmentation accuracy and completeness.

Conclusions:

• SEE provides a robust and effective solution for Incompletely-Supervised Concealed Object Segmentation.
• The method offers a plug-and-play capability to enhance existing segmentation models.
• This work advances the field of segmentation for challenging, real-world scenarios.