MoCIV: Modular Construction Installation Video dataset

MoCIV (Modular Construction Installation Video) contains 73,240 frames (3840×2160, 1 fps) from five mid-rise modular construction sites in South Korea, covering 41 volumetric module crane installation cycles (33 complete). Three annotation layers share one frame index: crane and module boxes held between keyframes across the full pool, with frame-level worker boxes on a 14,720-frame detection subset; one of five activity labels (Idling, Hooking, Lifting, Positioning, Returning) on every frame; and installation-cycle records derived from these labels. Leave-one-site-out split files are provided for both the detection subset and the full pool. Files: MoCIV_bundle.zip (annotations, metadata, splits, tools, README; 54 files) and five per-site image archives images_site_A.tar … images_site_E.tar (73,240 JPEG, 15.42 GB). Unzip the bundle and extract the tars into its root; images_manifest.csv lists every image with its SHA-256 hash. Code used to build, verify, and validate the release: https://doi.org/10.5281/zenodo.22908924