HZNU-BuildStage-4 Dataset

The dataset comprises four core construction phases—foundation excavation, main structure construction, topping-out, and completion—authentically reflecting the primary challenges in practical engineering monitoring: minimal inter-class visual variance, easily fragmented low-level physical features, and background environments prone to inducing "shortcut learning." To rigorously verify the models' true robustness, this dataset adopts a strict independent splitting protocol: the homologous Validation Set A is used exclusively for dynamic training monitoring and optimal weight selection (via an early stopping mechanism), while the cross-source Test Set B, completely isolated in terms of geographical distribution and data sources, is strictly excluded from any training or parameter selection processes.

It is reserved solely to provide the most objective and rigorous benchmark for absolute cross-scenario generalization performance.