Multi-Granularity Denoising and Bidirectional Alignment for Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation (WSSS) models using Class Activation Maps (CAMs) show promise but require complex pseudo-label generation.
• Existing methods struggle with efficiency and generalization to multi-class objects due to noisy saliency maps.

Purpose of the Study:

• To develop an efficient end-to-end WSSS approach that overcomes limitations of CAM-based methods.
• To address noisy labels and improve generalization for multi-class object segmentation using readily available saliency maps.

Main Methods:

• Proposed an end-to-end multi-granularity denoising and bidirectional alignment (MDBA) model.
• Implemented online noise filtering and progressive noise detection for image-level and pixel-level noise.
• Utilized bidirectional alignment with simple-to-complex image synthesis and adversarial learning.

Main Results:

• Achieved a mean Intersection over Union (mIoU) of 69.5% on the PASCAL VOC 2012 validation set.
• Reached a mIoU of 70.2% on the PASCAL VOC 2012 test set.
• Demonstrated improved generalization to complex, multi-class images.

Conclusions:

• The MDBA model effectively alleviates noisy labels and enhances generalization in WSSS.
• This approach offers an efficient and effective alternative for semantic segmentation tasks.
• The proposed method shows strong performance on benchmark datasets, with code publicly available.