Potatoes Dataset

The dataset contains color images of potatoes captured on an industrial conveyor sorting line under real post-harvest conditions. It was created for developing and evaluating machine learning models for automatic potato quality classification using deep learning methods. Images were acquired using a Basler Ace series industrial camera positioned perpendicularly to the sorting belt to avoid perspective distortion. The dataset includes both cleaned (washed) and unwashed potatoes to represent real-world variability caused by soil type, harvesting conditions, and surface contamination. Additional non-potato objects such as stones, clods of earth, and plant residues were included to reflect the natural environment of sorting lines. Each object in the dataset was manually annotated and categorized into one of four classes:Good – healthy, undamaged potatoes, suitable for market saleDamaged – potatoes with visible mechanical or biological defects (cut, rotten, or diseased)Plant – plant residues, potato tops, weeds, or leavesStone – stones or soil clods visually similar to potatoes The dataset is divided into the following subsets:Training set – mixed images of cleaned and unwashed potatoesValidation set – mixed samples used for model tuningTest sets – three independent parts:Cleaned test set – only cleaned potatoesUnwashed test set – only unwashed potatoesCombined test set – a mixture of both domains The metadata description in txt format is included in the attached zip file.