CS-7641-ML-PROJECT/autoencoder_training_data

Roadscape YOLO Road Surface Segmentation

Road surface segmentation of NAIP aerial imagery (0.3 to 0.6 m) in the southern United States, with three classes: paved, unpaved, and path. The labels are pseudo-labels: OpenStreetMap roads and their surface tags prompt SAM 3, and the resulting masks are converted to YOLO polygons.

4,106 tiles of 1024 px are split into 512 px windows, and windows without road pixels or with more than 10% unknown-surface road pixels are removed. Consequently, 10,679 windows from 3,636 tiles remain: 8,359 for training and 2,320 for validation, split by 10 km spatial block.

images/{train,val}/*.jpg   RGB windows, 512 x 512 px
labels/{train,val}/*.txt   YOLO polygons: class x1 y1 x2 y2 ... (0 paved, 1 unpaved, 2 path)
masks/{train,val}/*.png    SAM 3 reference masks: 0 background, 1 paved, 2 unpaved, 3 unknown, 4 path
data.yaml                  Ultralytics dataset configuration

File names are <stratum>_<tile x>_<tile y>_<row>_<col>, where row and col are the pixel offset of the window in its tile.

Imagery: USDA NAIP, public domain. Labels: derived from OpenStreetMap, © OpenStreetMap contributors, ODbL.