Coarse Mask Guided Interactive Object Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Interactive object segmentation refines object masks using user input like clicks.
• Current methods often lose click information through convolutional layers, hindering mask prediction accuracy.
• Deep learning approaches for interactive segmentation are gaining significant research interest.

Purpose of the Study:

• To develop a more efficient and accurate interactive object segmentation model.
• To address the information loss issue in click-based segmentation methods.
• To improve the exploitation of user interaction cues in deep learning models.

Main Methods:

• Propose a coarse mask guided (CMG) model for interactive object segmentation.
• Utilize a coarse module to encode clicks as query features and enrich them with backbone features via transformer layers.
• Generate coarse masks from enriched features to guide the main decoder for precise mask prediction.

Main Results:

• The CMG model effectively guides object mask prediction using coarse masks.
• The transformer-based coarse module and decoder are lightweight and computationally efficient.
• Achieved new state-of-the-art results on several segmentation benchmarks.

Conclusions:

• The proposed CMG model offers an effective and efficient solution for interactive object segmentation.
• Guiding mask prediction with coarse masks significantly improves performance.
• The method enhances the interaction process, making it smoother and more accurate.