Urban Features - Instance Segmentation (Tocantins)

This dataset consists of manually annotated image tiles extracted from very-high-resolution (3 cm/pixel) UAV orthoimagery of the Paraíso do Tocantins urban area, located in the Brazilian Cerrado biome. The source images were partitioned into training (70%), validation (20%), and test (10%) sets prior to patch extraction, ensuring that contiguous crops from identical source images are isolated within the same subset. A 4×4 spatial subdivision was subsequently applied to each image to accommodate the computational memory constraints of standard deep learning models. The finalized dataset yields 3,792 image tiles containing 76,981 annotations formatted to the COCO ans YOLO instance segmentation standard. Classes: tree roof wall sidewalk ruin pool