SDM-SAM2

This dataset provides the source code and experimental configurations associated with the paper “Semantic Dynamic Boundary Enhancement for Referring Remote Sensing Image Segmentation: SDM-SAM2.” It is released to support the reproduction and verification of the method implementation, model training, performance evaluation, and ablation procedures reported in the paper. SDM-SAM2 addresses referring remote sensing image segmentation by using Segment Anything Model 2 (SAM2) as the segmentation foundation and a BEiT-3 multimodal encoder to predict pixel-level masks of ground objects specified by natural-language expressions.The package contains the implementation of SDM-SAM2, training and evaluation scripts, experimental configurations, environment specifications, and usage documentation.

The method contains three main components: a stage-aware text memory module that stores and retrieves historical text states during hierarchical visual encoding; a task-adaptive multi-layer query fusion module that integrates query states from different layers of the mask decoder; and a text-weighted Sobel boundary loss that provides differentiable boundary supervision on continuous mask probability maps. Independent switches and combined configurations are provided for the three components, supporting experiments with the base model, single modules, two-module combinations, and the complete model.