Learning Latent Stable Patterns for Image Understanding With Weak and Noisy Labels.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised image understanding relies on image-level labels, lacking specific object localization.
• Existing methods assume error-free labels, which is often unrealistic due to label noise from pretrained predictors.
• Noisy image labels pose a significant challenge for accurate segmentation tasks.

Purpose of the Study:

• To develop a novel algorithm for weakly supervised segmentation with noisy image labels.
• To address the limitations of existing methods that assume label perfection.
• To improve the accuracy and robustness of image segmentation under label uncertainty.

Main Methods:

• A semantic space is constructed by embedding image labels through graphlet (superpixel cluster) encoding.
• A generative model, latent stability analysis, is proposed to identify stable patterns in the semantic space despite label noise.
• Graphlet semantics are inferred from stable mid-level patterns, offering a more secure approach than direct label transfer.
• Superpixel semantics are determined via maximum majority voting of correlated graphlets.

Main Results:

• The proposed algorithm demonstrates impressive performance in weakly supervised segmentation tasks.
• The method effectively handles noisy image labels, outperforming direct label transfer.
• Experimental results validate the algorithm's robustness with both hand-crafted and deep learning-based image descriptors.

Conclusions:

• The latent stability analysis effectively discovers robust patterns from noisy image labels for improved segmentation.
• Inferring semantics from stable mid-level patterns is more accurate than using noisy image-level labels directly.
• The developed algorithm offers a significant advancement in weakly supervised image segmentation under label noise.