Online Attention Accumulation for Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object attention maps from image classifiers are crucial for weakly supervised semantic segmentation.
• Current methods often fail to localize entire objects, focusing only on discriminative parts.
• This limitation hinders the performance of weakly supervised segmentation.

Purpose of the Study:

• To develop a novel weakly supervised method for identifying entire object regions.
• To enhance the integral localization of objects beyond discriminative parts.
• To improve the performance of weakly supervised semantic segmentation.

Main Methods:

• Proposed an online attention accumulation (OAA) strategy using attention maps from different training phases.
• Developed an attention drop layer to enlarge attention movement range for comprehensive region mining.
• Utilized cumulative attention maps as pixel-level supervision for further refinement.

Main Results:

• The OAA strategy effectively mines more object regions, overcoming limitations of single-phase attention maps.
• Incorporating the attention drop layer explicitly enlarges attention movement, leading to more integral regions.
• Achieved a state-of-the-art mIoU score of 67.2% on the PASCAL VOC 2012 segmentation benchmark.

Conclusions:

• The proposed OAA strategy with an attention drop layer significantly improves weakly supervised semantic segmentation.
• The method effectively generates integral object localization maps by accumulating attention over training phases.
• This approach offers a flexible plug-and-play solution for classification networks, enhancing object discovery and segmentation accuracy.