Auxiliary Tasks Enhanced Dual-Affinity Learning for Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation (WSSS) typically uses class activation mapping (CAM) for localization.
• Existing methods often combine CAM with saliency maps using heuristic thresholding for pseudo-label generation.
• There's a need to better leverage inter-task correlations for improved WSSS.

Purpose of the Study:

• To propose AuxSegNet+, a novel weakly supervised auxiliary learning framework for semantic segmentation.
• To explore the rich information in saliency maps and the correlation between saliency detection and semantic segmentation.
• To improve semantic segmentation using saliency detection and multilabel image classification as auxiliary tasks.

Main Methods:

• Introduced AuxSegNet+, a framework utilizing saliency detection and multilabel image classification as auxiliary tasks.
• Developed a cross-task affinity learning mechanism to learn pixel-level affinities from saliency and segmentation feature maps.
• Implemented a cross-task dual-affinity learning module for enhancing task-specific features and predictions via global context aggregation.

Main Results:

• Achieved new state-of-the-art results on weakly supervised semantic segmentation benchmarks (PASCAL VOC, MS COCO).
• Demonstrated the effectiveness of leveraging auxiliary tasks and cross-task affinity learning for improved segmentation.
• Showcased iterative performance improvement through cross-task affinity learning and pseudo-label updating.

Conclusions:

• AuxSegNet+ effectively utilizes auxiliary tasks and cross-task affinity learning for superior WSSS.
• The proposed cross-task dual-affinity learning module enhances feature representation and prediction accuracy.
• The framework offers a promising direction for advancing weakly supervised learning in computer vision.