Activation extending based on long-range dependencies for weakly supervised semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation (WSSS) commonly uses class activation maps (CAM) for pseudo-label generation, but CAM suffers from false and local activations due to limited annotation data.
• The high cost of detailed annotations necessitates efficient WSSS methods to improve segmentation accuracy.

Purpose of the Study:

• To propose a novel WSSS architecture for mining semantic information by modeling long-range dependencies to extend object masks.
• To enhance the reliability of pseudo-labels and improve the capture of superior semantic details in segmentation tasks.

Main Methods:

• A novel architecture is proposed to mine semantic information by modeling in-sample and inter-sample long-range dependencies.
• Images are divided into blocks, and self-attention is applied with fewer classes to capture long-range dependencies and reduce false predictions.
• Global to local weighted self-supervised contrastive learning is performed among image blocks to transfer local CAM activation to foreground areas.

Main Results:

• The proposed modules effectively capture superior semantic details and generate more reliable pseudo-labels.
• Experiments on PASCAL VOC 2012 demonstrated the model achieves 76.6% mIoU on the validation set and 77.4% mIoU on the test set.
• The model's performance surpasses that of comparison baselines, indicating its effectiveness in WSSS.

Conclusions:

• The developed method successfully addresses the limitations of traditional CAM in WSSS by leveraging long-range dependencies and contrastive learning.
• The proposed architecture provides a more robust approach to semantic information mining, leading to improved object mask generation and segmentation accuracy.