Junjue-Wang/Rank1-Ali-Tianchi-Real-World-Image-Forgery-Localization-Challenge

Combines Lovász loss and pixel-level Online Hard Example Mining to address extreme class imbalance in forgery localization, while employing multi-scale augmentation (1.0-3.0× zoom with 768×1024 progressive cropping) to amplify discriminative features between tampered and pristine regions. Employs an ensemble of ConvNeXt, Swin Transformer, and SegFormer encoders with UPerNet decoders, augmented with pseudo-labeling across multiple training iterations to expand the 40K dataset 2.7× and improve generalization.

199 stars. No commits in the last 6 months.