VISOR-NET: Visibility Estimation Based on Deep Ordinal Relative Learning under Discrete-Level Labels.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Visibility estimation is crucial for autonomous systems and environmental monitoring.
• Existing datasets often lack real-world complexity or continuous visibility labels.
• Deep learning methods require robust training data for accurate performance.

Purpose of the Study:

• To propose a novel end-to-end pipeline, VISOR-NET, for accurate visibility estimation.
• To address the limitations of current datasets by introducing new real-world and synthetic datasets.
• To evaluate the effectiveness of the proposed method against existing deep learning approaches.

Main Methods:

• Developed VISOR-NET, an end-to-end pipeline utilizing ordinal image information and relative relationships.
• Encoded ordinal information into ordered image pairs to learn a global ranking function.
• Collected a large-scale Foggy Highway Visibility Images (FHVI) dataset from real surveillance scenes.
• Synthesized an INDoor Foggy images dataset (INDF) with continuous annotations.

Main Results:

• VISOR-NET demonstrated superior performance in estimation accuracy, convergence rate, and model stability.
• The method was validated on public datasets, the new FHVI dataset (classification), and the INDF dataset (regression).
• Achieved effective inter-level to intra-level visibility estimation and approximate regression with discrete labels.

Conclusions:

• VISOR-NET offers a robust and effective solution for visibility estimation, outperforming existing methods.
• The introduction of FHVI and INDF datasets advances research in foggy image analysis.
• The proposed approach shows versatility, enabling both classification and regression tasks with improved accuracy and data efficiency.