SENTRY-CD: a synthetic data generation framework for military scene change detection
Abstract
Matched before and after defence imagery is difficult to obtain, and many public remote sensing benchmarks focus on large, long-term structural changes rather than short-term changes involving small mobile targets. We present SENTRY-CD, a synthetic data generation framework combining open-vocabulary grounding, promptable segmentation, latent-diffusion inpainting, deterministic object transformation, and context-aware weather or terrain editing in an inspectable ComfyUI workflow. Because object edits are executed through known masks and recorded transforms, annotations are derived from the edit itself rather than drawn manually; the same edit record supports box-level supervision for small-target change detection and dense masks for binary change segmentation. Scene-level condition editing and object-level semantic editing are controlled separately, allowing environmental appearance to be varied while the semantic edit is held fixed. A quality-screening stage removes pairs with incoherent changes, failed background reconstruction, or physically implausible placements. Across 4,095 training and 2,880 disjoint evaluation pairs spanning eight scene categories, a downstream change-segmentation probe reaches 0.898 mean per-image macro Dice and 0.859 mean per-image macro IoU. Performance is strongly left-skewed with a difficult failure tail; fog and snow produce the largest degradation. These results support annotation recoverability and graded synthetic difficulty, but do not by themselves establish photorealism or synthetic-to-real transfer.