Coarse-to-Fine Semantic Segmentation From Image-Level Labels.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep neural networks for semantic segmentation typically require extensive pixel-wise annotations, which are costly and time-consuming.
• Existing methods using object-level or image-level labels have limitations, especially with multi-category images.

Purpose of the Study:

• To propose a novel recursive coarse-to-fine semantic segmentation framework that utilizes only image-level category labels.
• To reduce the reliance on expensive pixel-wise annotations in training deep learning models for semantic segmentation.

Main Methods:

• An initial coarse mask is generated using a convolutional neural network (CNN)-based unsupervised foreground segmentation model.
• The coarse mask is enhanced by a graph model and then recursively refined by a fully convolutional neural network (FCNN).
• The framework handles multi-category objects with a single label per image, unlike other image-level methods.

Main Results:

• The proposed framework achieves performance comparable to state-of-the-art image-level label-based methods on the PASCAL VOC dataset, despite being trained solely on ImageNet.
• The framework demonstrates strong performance on the foreground object segmentation task, rivaling supervised methods on the Internet Object dataset.

Conclusions:

• The recursive coarse-to-fine framework offers an effective and efficient approach to semantic segmentation using minimal image-level labels.
• This method significantly lowers annotation requirements and broadens applicability to complex, multi-category scenes and foreground object segmentation.