Controllable Unsupervised Snow Synthesis by Latent Style Space Manipulation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Vehicle Technology

Background:

• Intelligent vehicle perception algorithms require diverse training data, often difficult and dangerous to acquire.
• Current unpaired image-to-image translation methods lack control over generation and multiple outputs.

Purpose of the Study:

• To develop a controllable image-to-image translation model for synthesizing challenging driving scenarios.
• To enhance the robustness of intelligent vehicle perception algorithms.

Main Methods:

• Proposed a generative adversarial network (GAN) with style and content encoders.
• Utilized a decoder for image reconstruction from encoded features.
• Implemented a self-regression module to constrain the style latent space for controllable generation.

Main Results:

• Successfully generated snow scenes on Cityscapes and EuroCity Persons datasets.
• Demonstrated controllable output generation by modifying hyperparameters and style codes.
• Validated the model's effectiveness in synthesizing realistic and varied image domains.

Conclusions:

• The proposed GAN model effectively synthesizes diverse and challenging visual data for intelligent vehicles.
• This approach reduces reliance on real-world data acquisition, accelerating algorithm development.
• The controllable generation capability offers significant advantages for robust perception system training.