Leveraging Geometric Structure for Label-Efficient Semi-Supervised Scene Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Segmentation

Background:

• Label-efficient scene segmentation seeks accurate per-pixel classification with minimal labeling.
• Current methods often overlook 3D geometric scene structures, limiting performance.
• Leveraging unlabeled data via consistency regularization or pseudo-labeling is common but can be improved.

Purpose of the Study:

• To introduce a novel Geometric Structure Refinement (GSR) framework for enhancing semi-supervised scene segmentation.
• To exploit 3D geometric scene structures for improved segmentation detail discrimination.
• To investigate optimal labeling strategies for semi-supervised segmentation tasks.

Main Methods:

• GSR framework utilizes unsupervised 3D reconstruction to calibrate dense pseudo-labels.
• Initial pseudo-labels are generated from fast, coarse annotations.
• The calibrated pseudo-groundtruth is used to train existing segmentation models without architectural changes or increased annotation costs.
• Exploration of mixed fine- and coarse-labeling strategies.

Main Results:

• GSR effectively enhances existing segmentation models (e.g., PSPNet, DeepLabv3+) with reduced annotations.
• Experiments on Cityscapes and KITTI datasets validate the framework's performance.
• Achieves 99% of fully supervised accuracy with only half the annotation effort.

Conclusions:

• GSR provides a flexible and effective method for label-efficient scene segmentation.
• Exploiting 3D geometric information significantly improves segmentation accuracy.
• Hybrid labeling strategies (fine + coarse) outperform traditional dense-fine annotations.