omorros/kitchen-fruit-vegetable-detection

Kitchen Fruit & Vegetable Detection

A real-world fruit and vegetable object detection dataset: 120 high-resolution photos taken in a home kitchen, with 585 hand-drawn bounding boxes across 14 classes. Every photo is tagged with its location, difficulty and camera angle, and labels come in both COCO and YOLO format.

Use it to test how a detector, or a vision-language model (VLM) such as GPT, Gemini or Claude, copes with fruit and vegetables outside the lab: cluttered counters, fridge drawers, grocery bags, mixed lighting, and look-alike pairs like lemon and orange.

Images	120 (4032x3024 or 3024x4032 JPEG, no EXIF or location data)
Boxes	585, hand-labelled
Classes	14: apple, banana, green bell pepper, red bell pepper, carrot, cucumber, grape, lemon, onion, orange, peach, potato, strawberry, tomato
Formats	Hugging Face datasets, COCO JSON, YOLO txt
Split	test only (an evaluation set)
License	CC BY 4.0

Why this dataset

Most fruit and vegetable datasets are studio photos of a single item on a white background. Models trained on them score close to 99% and then fall apart in a real kitchen. In the dissertation this set was built for, an EfficientNet classifier reached 99.75% on studio images, yet the YOLO + EfficientNet pipeline built on it scored only F1 0.55 on these photos. This set captures what the studio sets miss:

• Real surfaces and lighting: kitchen counter, fridge or shelf, wooden table, grocery bag or basket, chopping board or plate.
• Several items per photo: 2 to 9, sometimes touching.
• Look-alike pairs on purpose: lemon/orange, peach/orange, tomato/apple, red pepper/tomato, potato/onion, cucumber/green pepper.
• A balanced design you can slice results by.

Use cases

• Benchmark YOLO, DETR, RT-DETR or other object detectors on real-world produce.
• Evaluate vision-language models (GPT, Gemini, Claude, Qwen-VL and others) on counting and identifying food in a photo.
• Test open-vocabulary detectors (Grounding DINO, OWL-ViT, YOLO-World) with zero-shot prompts.
• Measure robustness to blur, noise and compression (see Degraded versions below).
• Sanity-check models for smart fridges, pantry and grocery inventory apps, and food waste tools.

Dataset design

Every photo was planned in advance so the set is balanced:

Factor	Levels	Photos per level
Location	kitchen counter, fridge or shelf, wooden table, grocery bag or basket, chopping board or plate	24
Difficulty	simple (2 to 3 items), medium (4 to 5), hard (6 to 8, some touching)	40
Camera angle	top-down, 45-degree, side	40

All items are unpackaged. Every photo was taken on the same phone on the same day, at full resolution. Boxes were drawn as polygons in Roboflow and converted to axis-aligned boxes.

Labelling rules

• One box per individual item (two apples means two boxes).
• Grapes: one box per bunch, not per berry.
• Strawberries: one box per berry.
• Bananas: one box per bunch if joined, one each if separated.

Classes

ID	Class	Boxes	Images
0	apple	54	50
1	banana	42	41
2	bell_pepper_green	29	29
3	bell_pepper_red	35	35
4	carrot	45	40
5	cucumber	34	34
6	grape	42	40
7	lemon	44	42
8	onion	43	39
9	orange	50	43
10	peach	43	41
11	potato	39	35
12	strawberry	37	33
13	tomato	48	45

Usage

Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("omorros/kitchen-fruit-vegetable-detection", split="test")
example = ds[0]
example["image"]                       # PIL image
example["location"], example["tier"], example["angle"]
example["objects"]["bbox"]             # [x_min, y_min, width, height] in pixels
example["objects"]["category_names"]   # e.g. ["banana", "apple"]

COCO

coco/annotations.json is a standard COCO detection file (category IDs 0 to 13, each tagged fruit or vegetable). It works with pycocotools, Detectron2, MMDetection and similar tools.

YOLO (Ultralytics)

from ultralytics import YOLO

model = YOLO("your_model.pt")
model.val(data="yolo/data.yaml", split="test")

Files

data/test/
  IMG_001.jpg ... IMG_120.jpg
  metadata.jsonl      one line per image: location, tier, angle, look-alike pairs, boxes
coco/
  annotations.json    COCO detection format
yolo/
  labels/IMG_*.txt    YOLO format (class cx cy w h, normalised)
  data.yaml
classes.txt           class names in ID order
preview.jpg

Degraded versions

To test robustness, the original study generated three degraded copies of every photo: Gaussian blur (kernel 15, sigma 6), Gaussian noise (sigma 50) and JPEG compression at quality 5. They are not stored here; the script is research/02-pipeline-comparison/evaluation/generate_degradations.py in the SnapShelf repo.

Baseline results

Count-based F1 from the dissertation (higher is better):

Condition	VLM (GPT-5.2)	YOLOv8s, 14 classes	YOLOv8s + EfficientNetB0
Clean	0.900	0.603	0.547
Blur	0.895	0.655	0.495
Noise	0.878	0.253	0.367
JPEG	0.866	0.512	0.517

All of the VLM's false positives on clean photos are grapes: it counted berries while the labels count bunches. Leaving grapes out, it scores F1 0.997. Gemini 3.1 Pro (F1 0.904) and Claude Opus 4.6 (0.867) were also tested on the clean set.

Origin

Built by Oriol Morros Vilaseca as the test set for a BSc (Hons) Software Engineering dissertation at Anglia Ruskin University, which compared VLM, YOLO and YOLO + CNN pipelines for a food inventory app. The app, SnapShelf, uses the winning approach to help people track what is in their kitchen and waste less food.

Limitations

• 120 photos, one home, one phone, one photographer. Treat it as a hard real-world check, not a large benchmark.
• 14 fruit and vegetable classes only. No packaged food.
• The grape rule (one box per bunch) penalises models that count individual berries.

License

CC BY 4.0. Use it for anything, including commercially, with credit.

Citation

@misc{morros2026kitchenproduce,
  author       = {Morros Vilaseca, Oriol},
  title        = {Kitchen Fruit \& Vegetable Detection: a real-world test set for food recognition},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/omorros/kitchen-fruit-vegetable-detection}},
  note         = {Test set for a BSc (Hons) Software Engineering dissertation, Anglia Ruskin University}
}