Regularized Loss With Hyperparameter Estimation for Weakly Supervised Single Class Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image segmentation typically requires precise pixel-level annotations, which are labor-intensive and costly.
• Existing weakly supervised methods often struggle with complex object properties and lack generalizability across diverse datasets.

Purpose of the Study:

• To develop an image-level weakly supervised segmentation approach that bypasses the need for pixel-precise annotations.
• To introduce a novel regularized loss function based on Conditional Random Field (CRF) modeling for guiding Convolutional Neural Networks (CNNs).
• To enable a unified and adaptable segmentation approach applicable to various tasks and datasets.

Main Methods:

• A regularized loss function inspired by classical Conditional Random Field (CRF) modeling is proposed.
• An annealing algorithm is developed to facilitate the training of CNNs with the regularized loss.
• A method for hyperparameter setting is introduced, addressing the challenge of lacking pixel-precise ground truth.

Main Results:

• The proposed method successfully guides CNNs towards accurate object segments without pixel-level supervision.
• State-of-the-art results were achieved in salient object segmentation, co-segmentation, and multi-class semantic segmentation tasks.
• The approach demonstrated effectiveness across different segmentation tasks and datasets using a standard CNN architecture.

Conclusions:

• The developed image-level weakly supervised segmentation method offers an efficient alternative to pixel-level annotation.
• The CRF-inspired regularized loss function and annealing algorithm provide a robust and generalizable solution for segmentation.
• This work advances the field of weakly supervised learning in computer vision, achieving competitive performance with reduced annotation effort.