Panoptic blind image inpainting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Accurate scene understanding is crucial for autonomous driving, but environmental factors like lens obstructions and weather degrade image quality.
• Existing image restoration methods often require ground truth data, which is unavailable in real-world autonomous driving scenarios.

Purpose of the Study:

• To develop a blind image inpainting model capable of accurately reconstructing corrupted images in real-world autonomous driving environments without ground truth.
• To enhance the reliability of scene understanding for autonomous vehicles by addressing image degradation issues.

Main Methods:

• Proposed a novel blind image inpainting model incorporating a panoptic map for detailed content representation.
• Designed an encoder-decoder architecture to predict both the panoptic map and the corrupted region mask.
• Implemented a mask refinement process to enhance the accuracy of map prediction and subsequent image restoration.

Main Results:

• The proposed model demonstrated superior performance compared to existing blind image inpainting methods on the Cityscapes and COCO datasets, measured by L1/L2 losses, PSNR, and SSIM.
• Achieved comparable results to image inpainting techniques that utilize additional information, highlighting its effectiveness without ground truth.

Conclusions:

• The developed blind image inpainting model effectively reconstructs corrupted images in challenging real-world conditions for autonomous driving.
• The approach significantly improves scene understanding accuracy by overcoming limitations of traditional inpainting methods.