Click any image to see its full annotation. Each sub-dataset carries different label types — the legend below groups them accordingly.
CSConstructionSite — caption + safety-rule violations + 3 object classes
Rule violation Excavator Rebar Worker (white hard hat)
SODA-VOCPascal VOC bounding boxes — 12 object classes, no scene caption
Person Helmet Vest Scaffold Hook Hopper Fence Electrical box Board Handcart Rebar Slogan
SODA-KTSHCaption-only sub-dataset — 1–5 paraphrased scene captions per image, no bounding boxes