Incorporating Network Built-in Priors in Weakly-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Pixel-level annotations for semantic segmentation are costly and time-consuming.
• Weak supervision using image tags offers a more efficient alternative but often suffers from poor localization accuracy.
• Existing methods using objectness priors require additional annotations or yield inaccurate boundaries.

Purpose of the Study:

• To develop a novel method for accurate mask extraction from pre-trained networks for weakly supervised semantic segmentation.
• To eliminate the need for external objectness modules and pixel-level or bounding-box annotations.
• To improve localization accuracy in tag-based semantic segmentation tasks.

Main Methods:

• Extracting foreground/background masks directly from the activations of higher-level convolutional layers in pre-trained networks.
• Generating multi-class masks by fusing foreground/background masks with information from a weakly-supervised localization network.
• Utilizing these extracted masks in conjunction with a weakly-supervised training loss.

Main Results:

• The proposed method successfully extracts accurate masks without external objectness modules.
• Fusion of foreground/background masks with localization network information yields effective multi-class masks.
• The approach achieves state-of-the-art results in tag-based weakly supervised semantic segmentation.

Conclusions:

• The novel method effectively leverages pre-trained networks for accurate mask generation in semantic segmentation.
• This approach significantly enhances localization accuracy in weakly supervised settings.
• It represents a state-of-the-art solution for tag-based weakly supervised semantic segmentation, reducing annotation dependency.