akshara-ns/dessert-image

Dessert Image Dataset

Purpose

Photographs of desserts labelled by whether the dessert is chocolate-based. Built for HW1 of Designing and Prototyping AI Systems.

Composition

Each sample is one 224x224 RGB image with the binary target is_chocolate.

Provenance field	Description
source_id	Unique id for the row
group_id	Related samples share a group; the split is made by group
parent_id	The original this row came from (itself, for originals)
second_parent_id	Second parent, for methods that blend two rows
mix_weight	Blend weight; 1.0 unless two parents were mixed
augmentation	Which method produced the row (none for originals)
augmentation_detail	The exact parameters applied
is_augmented	False for collected samples, True for synthetic ones
collected_by	Author of the original sample

Collection

All photos were taken by the author on her own phone, in cafes, restaurants and at home. Photos containing people or bystanders were removed before labelling.

Preprocessing and augmentation

Each photo was scaled to fit inside 224x224 and padded with white on the short side, preserving the aspect ratio so nothing is cropped away; re-saving also removes EXIF/GPS metadata. Synthetic images were generated from training originals only, by five named methods: horizontal flip, rotation (+/-15 degrees, white fill), crop of 80-98% then resize, brightness/contrast within +/-15%, and a combined variant. Hue and colour-channel shifts were deliberately avoided, since the target is judged largely by colour. Duplicate and unchanged outputs were removed.

Method	Rows
combined	122
crop_resize	125
horizontal_flip	25
none	25
photometric	125
rotation	125

Labels

is_chocolate was assigned by the author from first-hand knowledge of each dessert. Single-annotator labels; borderline cases (a cocoa-dusted tiramisu, a small chocolate garnish) reflect her judgment.

Splits

Split	Original rows	Synthetic rows	Total
train	25	522	547
validation	5	0	5
test	5	0	5

Split 70/15/15 with random seed 24679, stratified by is_chocolate and grouped by group_id, before any augmentation. Synthetic rows derive only from training originals, so validation and test remain untouched. Filter is_augmented == False to recover exactly the collected samples.

Intended use and limits

Intended: coursework and prototyping: small-scale image classification, and studying how image augmentation affects a small training set.

Not intended for: production food recognition, or nutrition and allergen decisions.

The data reflects one person's desserts, habits and cities, so it will not generalize broadly. Augmentation adds variety but no new information: every synthetic row is a transformation of one (or two) of the 25 training originals. Validation and test are small (5 and 5 samples), so metrics from them are noisy indicators rather than precise scores.

Ethical considerations

No faces, people or identifying information appear in any image; photos containing bystanders were excluded before labelling. GPS metadata was removed during resizing.

License

CC BY 4.0. Attribution required.

Acknowledgments

Collaboration: this dataset was collected, labelled and written by the author alone. No classmate contributed data, labels or text. Discussion of the assignment with classmates, where it happened, did not extend to sharing samples.

Tools: Google Colab, Hugging Face datasets and huggingface_hub, pandas, scikit-learn, matplotlib, Pillow. Claude (via Claude Code) assisted with the notebook code, as described below.

AI usage disclosure

All original data is the author's own work. No generative AI was used to create any data. The notebook code was written with AI assistance (Claude, via Claude Code) and reviewed and run by the author. Augmentation is rule-based and seeded, not generative.