A Robust Real-world Change Detection Transformer

Change detection aims to localize changed objects in an image compared to a reference image. To effectively address both pixel-level change detection in remote sensing images and object-level change detection in natural images within a unified pairwise-interaction framework, we present the Change DEtection TRansformer (Change DETR), a robust query-based model tailored for real-world change detection. Building on the advanced DETR, Change DETR comprises a backbone, a transformer encoder-decoder, and crossover attention modules for inter-image interactions. The adaptability of query tokens, configurable for object-level or semantic-level concepts depending on task settings, allows for the decoding of both object-level and pixel-level predictions. Moreover, Change DETR employs crossover attention layers in both the encoder (EnCA) for iterative context feature interactions and the decoder (DeCA) for sequential interactions between query tokens and context features. The synergistic integration of EnCA and DeCA, referred to as the Dual Crossover Attention module (DuCA), significantly improves Change DETR's representation capabilities for change detection tasks. Furthermore, recognizing the limitations of existing benchmark datasets, which are mostly synthetic or constrained to identity transformations and do not adequately reflect real-world conditions, we develop the Real-Change dataset. This dataset features 3D geometric transformations from natural scenes, providing a more realistic evaluation. Our method achieves the best performance across five natural scene change detection datasets, achieving 75.47% and 82.36% in Average Precision (AP) on the COCO-inpainted and Kubric-Change benchmarks, respectively. It also shows impressive robustness in pixel-level remote sensing change detection, attaining top results on two widely recognized benchmarks and proving its robustness across diverse image domains.