Diamant Potato Dataset for ROI and Surface Defect Instance Segmentation

Two annotated image datasets of Solanum tuberosum cv. Diamant (Diamant Potato) supporting a cascaded quality-inspection pipeline for export-grade potato sorting. Stage 1 isolates potatoes from the background; Stage 2 segments surface defects on the resulting crops. All images were collected from three local markets in Dhaka, Bangladesh (Ashulia, Banasree, Khilgaon) to capture variation across commercial supply. Both datasets were collected and annotated for this work. 1. ROI dataset (roi_dataset/) •Images: 2,386 total, containing 2,581 annotated potato instances. •Composition: 1,938 single-potato images (each tuber photographed from front, back, left and right), 304 group images with multiple tubers, 74 pure distractor images, 70 mixed distractor images. •Distractor design: pure distractors (ginger roots, eggs and plain brown surfaces) suppress false positives on frames with no potato; mixed distractors force discrimination within a single frame. Pure distractor images carry empty label files as intentional negative samples. •Classes: one (0: Potato). •Splits: pre-partitioned 70/20/10 into 1,670 training, 477 validation and 239 test images. •File format: JPG at 3000×3000 pixels. 2. Defect dataset (defect_dataset/) •Images: 1,724, containing 6,576 annotated defect instances. •Classes: four (0: Scab, 1: Rot, 2: Greening, 3: Mechanical Damage). •Instances by class (total / train / val / test): Scab 4,390 / 3,188 / 548 / 654; Rot 1,268 / 903 / 175 / 190; Mechanical Damage 793 / 562 / 108 / 123; Greening 125 / 79 / 17 / 29. •Splits: pre-partitioned 75/10/15 into 1,281 training, 176 validation and 267 test images, stratified by class. •Preprocessing: all images are Green-Boosted. Pixels in the green HSV band (OpenCV hue 35–85, saturation ≥40, value ≥40) have saturation multiplied by 1.5 and value by 1.3, clipped to 255; background pixels remain black. The same transform must be applied at inference. •File format: lossless PNG, background-removed crops on uniform black. Class definitions: •Scab (common scab, powdery scab, elephant hide): raised, corky, crater-like tan or brown lesions with intact underlying skin. •Rot (late blight, dry rot, bacterial soft rot): collapsing, sunken, water-soaked, or dark brown to black tissue decay. •Greening (chlorophyll accumulation, solanine discolouration): superficial greenish discolouration from light exposure. •Mechanical Damage (harvest gouges, growth cracks, skinning, dry bruises): sharp cuts or abrasions exposing dry, clean inner flesh. Interpretation and use: The two sets are designed to be used in sequence. They can also be used independently, the ROI set for background-robust produce detection and the defect set for surface defect segmentation on isolated tubers. The defect class distribution is heavily skewed, Scab accounts for 67% of instances and Greening for 1.9%. Both partitions were made at the image level.