wgcban/ChangeFormer

Implements a dual-branch vision transformer architecture with cross-attention mechanisms for bitemporal remote sensing image analysis, enabling multi-scale training and inference for change detection tasks. Built on PyTorch with support for standard segmentation datasets (LEVIR-CD, DSIFN-CD) and compatible with ADE20k pretrained initialization for accelerated convergence. Provides end-to-end training pipelines with configurable loss functions (cross-entropy, focal loss, mIoU) and optimization strategies (AdamW, SGD).

581 stars. No commits in the last 6 months.