DLnet With Training Task Conversion Stream for Precise Semantic Segmentation in Actual Traffic Scene.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Semantic segmentation models often fail in real-world scenarios due to a performance gap.
• Existing training task conversion (TTC) and domain adaptation methods have limitations, hindering performance.
• Maintaining high performance during TTC is a significant challenge in computer vision.

Purpose of the Study:

• To propose a novel deep learning model, DLnet, for effective training task conversion (TTC).
• To bridge the performance gap between models trained on datasets and their application to real-world scenes.
• To enhance the robustness and performance of semantic segmentation models in actual environmental conditions.

Main Methods:

• Development of DLnet, a deep learning network with three key innovations.
• Experimental validation of DLnet on multiple benchmark datasets.
• Evaluation of DLnet's performance in diverse, real-world urban scenes.

Main Results:

• DLnet achieved state-of-the-art quantitative performance across four popular datasets.
• The model demonstrated outstanding qualitative performance in four distinct urban environments.
• DLnet exhibited robustness and strong performance, validating its effectiveness.

Conclusions:

• DLnet successfully addresses the performance gap in semantic segmentation for real-world applications.
• The proposed model offers a robust solution for training task conversion, maintaining high performance.
• While not real-time, DLnet provides acceptable moderate performance for practical use.