After a Decade: Bringing Shadow Removal into the Real World with Agentic Training Data

TL;DR

A new dataset called AgenticShadow was created to improve shadow removal in real-world images by using an innovative workflow that generates shadow-free targets. This approach significantly reduces color distribution differences and enhances model performance across various domains.

Key Takeaways

• Utilize the AgenticShadow dataset for better shadow removal results.
• Implement the agentic workflow for generating shadow-free targets.
• Focus on physics-based generation techniques in your models.
• Evaluate model performance using LAB RMSE metrics.