Cross-Weather Image Alignment via Latent Generative Model with Intensity Consistency.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Image alignment is crucial for many vision tasks but challenging across different domains like varying weather and seasons.
• Existing methods struggle with cross-weather/season road scene alignment.

Purpose of the Study:

• To propose a novel method for cross-weather road scene alignment.
• To address the limitations of current image alignment techniques in diverse environmental conditions.

Main Methods:

• Formulating alignment as a constrained 2D flow optimization problem with latent encoding.
• Utilizing a pre-trained generative adversarial network (GAN) to parameterize an image manifold.
• Decoding latent encoding into an intensity-constancy image on the learned manifold.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches on cross-weather/season road scene datasets.
• Demonstrated realistic warping effects by constraining the warped latent image to match the target image.

Conclusions:

• The latent generative model with intensity constancy offers a robust solution for cross-weather road scene alignment.
• This approach advances the field of image registration for challenging, real-world scenarios.