Learning from Weak and Noisy Labels for Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation (WSSS) utilizes image-level labels instead of pixel-level annotations.
• Leveraging publicly available image-level tags from platforms like Flickr enables large-scale applications.
• Existing WSSS methods often struggle with noisy and limited weak labels.

Purpose of the Study:

• To develop a WSSS method robust to noisy image-level labels.
• To address the challenge of learning segmentation models from both weak and noisy data.
• To propose a novel approach for label noise reduction in WSSS.

Main Methods:

• The WSSS problem is reframed as a label noise reduction task.
• Images are segmented into superpixels, and weak labels are propagated to these superpixels.
• A novel L1-optimisation based sparse learning model is formulated to detect and correct noisy superpixel labels.
• An efficient learning algorithm with an intermediate labelling variable is developed to solve the L1-optimisation problem.

Main Results:

• The proposed method achieves state-of-the-art performance on benchmark datasets with noise-free labels.
• The method significantly outperforms existing approaches when dealing with noisy weak labels.
• Demonstrated effectiveness in identifying and correcting noisy superpixel labels for improved segmentation.

Conclusions:

• The developed method offers a robust solution for WSSS, particularly in the presence of label noise.
• This approach enhances the utility of large-scale, weakly labelled image datasets.
• The L1-optimisation based sparse learning model provides an effective mechanism for noisy label detection and correction in WSSS.