FruitLife: A Longitudinal, Time-Stamped Image Dataset of Postharvest Fruit Spoilage Progression

FruitLife is a longitudinal, time-stamped image dataset for studying postharvest fruit spoilage. The main collection tracks 48 physical specimens across eight species (apple, apricot, banana, black plum, fig, grapes, nectarine, tomato) under three storage environments (sun, shade, indoor room), imaged three times a day from purchase to advanced spoilage (1,780 images). Every frame carries an in-frame colour-and-scale reference card, a ruler, and a manual ordinal spoilage stage (0-4). A secondary pilot subset (751 images, eight species, uncontrolled backgrounds) supports in-the-wild robustness testing. Because each specimen is imaged repeatedly along its own known trajectory, capture timestamps alone supply free temporal supervision, without additional annotation effort. The dataset is validated on four axes: intra-rater label reliability (weighted κ = 0.804 on a blind re-labelling sample), temporal consistency (zero monotonicity violations in the original labels), calibrated colorimetric and shrinkage trajectories (card-referenced CIELAB, ruler-referenced area), and ordinal spoilage-stage classification baselines across three generalisation protocols of increasing difficulty -- leave-one-species-out (unseen fruit species), leave-one-storage-out (unseen environment), and leave-one-specimen-out (unseen individual fruit, same species/storage) -- where a frozen self-supervised (DINOv2) backbone clearly outperforms a fully fine-tuned convolutional baseline on both out-of-distribution protocols despite training roughly 5,000x fewer parameters, while the fine-tuned CNN takes a small edge only on the easiest, in-distribution protocol. Known caveats (see the paper's Limitations section for full detail): labels come from a single primary annotator -- reported reliability above is intra-rater, not inter-rater; the colour reference card is hand-painted and uncertified, so colour correction should be read as relative (session-to-session) normalisation, not absolute colorimetry; "sun" storage bundles irradiance and heat with no environmental logger; and per-species/storage specimen counts remain modest (two replicates per cell in most of the design). New in v1.3: shadow-free fruit masks and measurements (calibration_v2.csv), validated against manual outlines on 60 images; blind-revised stage labels for the images bracketing each specimen's stage transitions (two raters blinded to storage, time and the original labels, with adjudication), with the individual ratings; per-specimen stage event intervals; pilot timestamp flags; and the full analysis code. Earlier descriptions stated that specimens were imaged twice daily; they were imaged three times a day (about 06:00, 13:00 and 17:00), as the timestamps show. No image or original label changed. New in v1.4: updated analysis code (fruitlife_code_v1.4.tar.gz), adding the code for the workflow figure of the companion kinetics paper and for its fruit-only stage prediction (src/data/fruit_only_crops.py: shadow-free fruit cut-outs at a common physical scale, with the colour card and ruler removed). No image, label or measurement changed. New in v1.5: analysis code and results of the companion decay-clock paper (fruitlife_code_v1.5.tar.gz, fruitlife_paper3_results_v1.5.tar.gz): fruit-only, scale-normalized crops, training and evaluation of the decay clock (a score trained with a within-specimen ranking loss on capture timestamps and an ordinal loss on stage labels), remaining-time prediction, the replication of a published avocado baseline on the public 'Hass' avocado dataset (manifest only), fold definitions, per-fold results and per-frame scores. No image, label or measurement changed. Supports shelf-life regression, ordinal spoilage staging, and self-supervised temporal-ordering research.