Relative CNN-RNN: Learning Relative Atmospheric Visibility From Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Atmospheric Science

Background:

• Estimating atmospheric visibility typically requires specialized sensors or weather data.
• Existing methods may be costly or limited in scope.

Purpose of the Study:

• To develop a deep learning approach for direct estimation of relative atmospheric visibility from outdoor photographs.
• To create a cost-effective and scalable method for visibility assessment.

Main Methods:

• A novel Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) coarse-to-fine model was developed.
• The model leverages a large dataset of internet images and human annotations for training.
• It utilizes shortcut connections to integrate global (CNN) and local (RNN) scene information.

Main Results:

• The CNN-RNN model effectively estimates relative atmospheric visibility directly from images.
• The approach demonstrates the potential for adapting relative visibility predictions to absolute values in certain contexts.
• A large-scale annotated dataset of approximately 40,000 images with 0.2 million annotations was created.

Conclusions:

• Deep learning offers a viable alternative for direct atmospheric visibility estimation.
• The proposed CNN-RNN model provides a robust and data-driven solution for visibility assessment.
• The publicly released dataset will facilitate future research in this domain.