Paper Reading (2024/04/10)

使用diffusion mode生成synthetic images，用于enhance traning data:

1. FreeMask: Synthetic Images with Dense Annotations Make Stronger Segmentation Models (NurIPS 2024)

Challenges: 

• Generating in-domain images
• Generating high-quality images
• Generating hard samples to improve model generalization ability

How to solve the challeges:

• FreestyleNet, which can generate high-fedility images conditioned on semantic masks
• Incorrectly synthesized regions or images will exhibit significant losses, if evaluated under a model pre-trained on real images ---- > if the loss of a synthetic pixel surpasses the average loss of its corresponding class by a certain margin, it will be marked as a noisy pixel and ignored during loss computation
• with recorded class-wise average losses, it can calculate the overall loss of a semantic mask, which can represent its global hardness. It will generate more samples for those hard semantic masks and otherwise the opposite

短语：“It is even on par with”

2. DiffuLT: How to Make Diffusion Model Useful for Long-tail Recognition