STC: A Simple to Complex Framework for Weakly-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep convolutional neural networks (DCNNs) have advanced semantic object segmentation.
• Training DCNNs typically requires extensive pixel-level annotations, which are expensive and time-consuming.
• Existing methods struggle with the cost and effort of generating detailed segmentation masks.

Purpose of the Study:

• To propose a novel framework for semantic segmentation that minimizes reliance on pixel-level annotations.
• To develop a cost-effective method for training DCNNs using only image-level annotations.
• To enhance the efficiency and scalability of semantic segmentation model training.

Main Methods:

• A Simple to Complex (STC) framework is proposed, progressing from simple to complex image datasets.
• Initial-DCNN is trained using automatically generated saliency maps from simple images.
• Enhanced-DCNN and Powerful-DCNN are trained iteratively using image-level annotations and progressively refined segmentation masks.

Main Results:

• The STC framework demonstrates superior performance on the PASCAL VOC 2012 segmentation benchmark.
• The proposed method achieves state-of-the-art results compared to existing approaches.
• Successful training of DCNNs for semantic segmentation using only image-level annotations was achieved.

Conclusions:

• The STC framework offers a viable and efficient alternative to traditional pixel-level annotation methods.
• This approach significantly reduces the financial and human resources required for semantic segmentation model development.
• The study highlights the potential of leveraging image-level annotations for robust semantic segmentation.