Improving Semantic Segmentation via Efficient Self-Training.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Semantic segmentation models, particularly Fully Convolutional Networks (FCNs), have advanced significantly.
• Training these deep learning models demands extensive pixel-wise annotations, which are costly and time-consuming.
• Data imbalance is a common challenge in semantic segmentation tasks.

Purpose of the Study:

• To develop a self-training framework to reduce reliance on expensive pixel-wise annotations for semantic segmentation.
• To address the data imbalance problem inherent in semantic segmentation datasets.
• To improve the efficiency and effectiveness of training deep learning models for semantic segmentation.

Main Methods:

• Introduced a self-training framework leveraging pseudo-labels generated from unlabeled data.
• Proposed a centroid sampling strategy to ensure uniform sample selection across all classes, mitigating data imbalance.
• Implemented a fast training schedule to reduce computational overhead, enabling the use of more pseudo-labels.

Main Results:

• The Centroid Sampling based Self-Training (CSST) framework achieved state-of-the-art performance on the Cityscapes and CamVid datasets.
• Models trained with CSST on the original PASCAL VOC 2012 dataset outperformed those trained on a larger augmented set, highlighting effectiveness with limited annotations.
• Demonstrated promising few-shot generalization capabilities across different datasets (Cityscapes to BDD100K and Mapillary).

Conclusions:

• CSST is an effective approach for semantic segmentation, significantly reducing the need for extensive manual annotations.
• The centroid sampling strategy successfully handles data imbalance, leading to more robust model training.
• The framework shows potential for real-world applications requiring efficient training and strong generalization with limited labeled data.