Multi-Target Pan-Class Intrinsic Relevance Driven Model for Improving Semantic Segmentation in Autonomous Driving.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current semantic segmentation models heavily depend on deep learning for feature extraction, achieving high performance but facing limitations in refining target segmentation.
• Existing methods often focus on intra-class and inter-class relationships, leaving room for improvement in cross-class target relevance.

Purpose of the Study:

• To propose a novel semantic segmentation model that refines target main body segmentation by utilizing multi-target pan-class intrinsic relevance.
• To address the performance limitations of current deep learning-based semantic segmentation networks.

Main Methods:

• Establishment of prior knowledge for multi-target pan-class intrinsic relevance (RPK-Est module).
• Extraction of pan-class intrinsic relevance features using a multi-target node graph and graph convolution network (RF-Ext module).
• Integration of intrinsic relevance and semantic features via generative adversarial learning at the gradient level (RF-Int module).

Main Results:

• The proposed model demonstrated outstanding performance in semantic segmentation tasks.
• The model achieved superior results compared to state-of-the-art models across four authoritative datasets.
• The integration strategy effectively utilized intrinsic relevance features for improved semantic segmentation.

Conclusions:

• The developed model successfully refines target main body segmentation by exploiting pan-class intrinsic relevance.
• The novel approach offers a significant advancement in semantic segmentation, overcoming limitations of existing deep learning methods.
• The findings highlight the potential of cross-class relevance for enhancing segmentation accuracy and robustness.