ParsaVictor/tomato-ripeness-detection

🍅 Unified Tomato Detection & Ripeness Dataset

8,497 images · 74,486 tomato boxes · 4 ripeness classes (unripe, half-ripe, ripe, rotten) · YOLO format · leakage-free holdouts

Six public tomato datasets merged into one consistent schema, de-duplicated and split for honest evaluation. Trained models: complete pipeline · detector · ripeness classifier. Used to train the models in ParsaVictor/tomato-ripeness-detection (detection, ripeness grading and conveyor counting with YOLO26 + ByteTrack).

Files

File	Content
tomato_unified_detection.zip	unified/ — 4 parts (u_defects, u_sxsc7, u_conveyor, u_extra), each with train/valid/(test) → images/ + YOLO labels/ and a data.yaml
tomato_ripeness_crops.zip	cls_tomato/{train,val,test}/{raw,half_ripe,ripe,rotten}/ — 160×160 crops for ripeness classification
conveyor_holdout.txt, sxsc7_holdout.txt	images reserved for testing (never train on them)
videos/tracking_counting_demo.mp4	demo: tracking IDs, counting line and ripeness dashboard
videos/conveyor_tracking.mp4	demo on real conveyor-belt frames

Classes

raw (unripe) · half_ripe · ripe · rotten · tomato (sources without ripeness labels: sxsc7, conveyor)

Class	Boxes
raw	31,406
half_ripe	16,952
ripe	15,307
rotten	1,575
tomato	9,246

What was done

• Class names from 6 sources mapped to one schema (e.g. fully-ripe → ripe, semi-ripe → half_ripe, Fresh → ripe, B/Tomato/Tomatob → tomato)
• Fixed holdouts: 43 conveyor frames and 57 sxsc7 images; 221 images too similar to a holdout image set aside
• 1,234 training images that were near-copies of test images removed (perceptual dHash)
• Ripeness crops: 50,776 train / 8,579 val / 768 test (rotten ×4 with flips/rotations in train)

Baseline results (YOLO26s + YOLO26s-cls)

Test set	Recall	Precision	Count accuracy
Tomato Defects test	99.0%	93.3%	93.8%
Conveyor holdout	84.9%	90.0%	89.7%
sxsc7 holdout	93.3%	94.8%	96.3%

Ripeness: 80.8% (4-class) · 98.6% (sound vs rotten).

Sources & license (all CC BY 4.0)

• Tomato Defects — Universitas Hasanuddin
• tomato-sxsc7 — Dai Tran Van
• tomato_count_project — canberk
• tomato-v1 — papapeto
• tomato-detection-sorting
• tomato-classification

This unified version: CC BY 4.0. Please credit the original authors and this dataset.

@misc{karkooti2026tomato,
  author = {Parsa Karkooti},
  title  = {Unified Tomato Detection and Ripeness Dataset},
  year   = {2026},
  url    = {https://github.com/ParsaVictor/tomato-ripeness-detection}
}