Memory-Based Cross-Image Contexts for Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation (WSSS) aims to reduce the need for pixel-level annotations by using weaker forms of supervision.
• Image-level labels are a common form of weak supervision in WSSS, but existing methods often overlook valuable cross-image contextual information.

Purpose of the Study:

• To develop an effective method for WSSS using only image-level labels.
• To exploit cross-image contexts to generate higher-quality pseudo-masks for improved segmentation performance.

Main Methods:

• An end-to-end cross-image context module was designed, incorporating a memory bank and a transformer-based cross-image attention mechanism.
• The memory bank stores extracted cross-image contexts from image feature encodings.
• The attention module utilizes memorized contexts to enhance pseudo-mask generation.

Main Results:

• Experiments on Pascal VOC 2012 and COCO datasets demonstrate the effectiveness of incorporating cross-image contexts.
• The proposed method achieved state-of-the-art performance in weakly supervised semantic segmentation tasks.
• The approach significantly improves the quality of pseudo-masks generated from image-level labels.

Conclusions:

• Cross-image contexts are crucial for improving pseudo-mask generation in WSSS.
• The developed cross-image context module offers a powerful solution for leveraging this information.
• This work advances the field of WSSS by achieving superior results with image-level supervision.