Beyond Group Splits: Specimen-Level Cross-Validation and Visual Attribution for Remaining-Shelf-Life Regression in Climacteric Fruit

Rovhona Mudau; Jean Frederic Isingizwe Nturambirwe; Clement Nthambazale Nyirenda

cs.CVOct 7, 2026

Beyond Group Splits: Specimen-Level Cross-Validation and Visual Attribution for Remaining-Shelf-Life Regression in Climacteric Fruit

Authors: Rovhona Mudau, Jean Frederic Isingizwe Nturambirwe, Clement Nthambazale Nyirenda

Organizations: Department of Computer Science University of the Western Cape Cape Town, South Africa · eResearch Office University of the Western Cape Cape Town, South Africa · Department of Computer Science / eResearch Office University of the Western Cape Cape Town, South Africa

Abstract

Estimating remaining shelf life (RSL) from images could provide affordable decision support for perishable produce, but evaluation protocols can substantially affect reported performance when repeated images are available from the same biological specimen. We use the Hass Avocado Ripening dataset, comprising 8,834 image-RSL pairs from 426 fruits across three storage regimes, to evaluate a frozen ImageNet-pretrained visual backbone with a lightweight regression head. Our contributions are threefold: we quantify the effect of observation-level versus specimen-disjoint evaluation, compare lightweight and heavier visual backbones under specimen-disjoint cross-validation, and examine their spatial attributions using Grad-CAM. Across ten observation-level random splits, the model achieves a mean RMSE of 2.37 days with a standard deviation of 0.03 days, whereas specimen-disjoint 5-fold cross-validation yields a mean RMSE of 3.12 days with a standard deviation of 0.11 days. The corresponding mean coefficient of determination is 0.553. A matched per-specimen comparison confirms higher error under specimen-disjoint evaluation, with a probability value below 0.001 across 426 specimens, showing that observation-level partitioning gives a substantially more optimistic estimate for this dataset and model configuration. Under specimen-disjoint evaluation, MobileNetV3-Small (0.93 million parameters) achieves accuracy comparable to ResNet-18 while providing substantially higher throughput, and Grad-CAM reveals differences in spatial attribution between the lightweight backbones. These results support specimen-disjoint evaluation and attribution analysis when assessing lightweight vision models for longitudinal shelf-life prediction.

Figures & tables

Reference	Task	Protocol category	Evaluation design
[ 13 ]	Hass avocado days-to-ripeness regression	Random sub-image split	Each hyperspectral fruit image’s sub-images were assigned to training, validation, and test sets; as a result, source-image identity was not disjoint across partitions.
[ 2 ]	Hass avocado ripening classification and shelf-life estimation	Specimen-level split	To prevent observations from the same sample from crossing partitions, avocado samples were uniquely assigned to 70/15/15 training, validation, and test subsets.
[ 14 ]	Hass avocado firmness prediction with RSL recommendation	Random image split	A total of 1,400 images were randomly shuffled into 80/10/10 training, validation, and test subsets; specimen grouping was not reported for the regression split.
[ 16 ]	Strawberry storage-age prediction	Condition-level split	Control-storage samples were used for model development and samples stored under modified-atmosphere packaging were used as the test condition.
[ 15 ]	Mango shelf-life-stage classification	Independent external validation	Internal 10-fold cross-validation was supplemented by evaluation on a separate online mango image dataset.
This study	Hass avocado RSL regression	Random, condition-level, and specimen-disjoint evaluation	Observation-level random splitting is used as a leakage diagnostic; storage regimes are held out for cross-condition analysis; primary within-dataset generalisation is assessed using specimen-disjoint 5-fold cross-validation.

TABLE I: Evaluation protocols in representative fruit ripening and shelf-life studies.

Protocol	RMSE img (d)	MAE img (d)	$R^{2}_{\mathrm{img}}$	RMSE fruit (d)	MAE fruit (d)
Random observation-level split	$2.37\pm 0.03$	$1.73\pm 0.03$	$0.744\pm 0.006$	$2.42\pm 0.04$	$1.76\pm 0.02$
Leave-one-storage-regime-out	5.18	4.07	-0.227	4.68	3.60
Specimen-disjoint 5-fold CV	$3.12\pm 0.11$	$2.35\pm 0.08$	$0.553\pm 0.023$	3.07	2.30

TABLE II: Image-level and fruit-balanced performance under the three evaluation protocols.

Backbone	Par. (M)	Thr.	RMSE (d)	MAE (d)	$R^{2}$
MobileNetV3-S	0.93	76/s	$\mathbf{2.97\!\pm\!0.14}$	$\mathbf{2.24\!\pm\!0.12}$	$\mathbf{0.596\!\pm\!0.027}$
MobileNetV2	2.22	—	$3.12\!\pm\!0.11$	$2.35\!\pm\!0.08$	$0.553\!\pm\!0.023$
ResNet-18	11.18	11/s	$3.09\!\pm\!0.08$	$2.35\!\pm\!0.07$	$0.562\!\pm\!0.018$

TABLE III: BACKBONE COMPARISON UNDER SPECIMEN-DISJOINT 5-FOLD CROSS-VALIDATION

Stage	MobileNetV2	MobileNetV3-Small
1 (fresh)	0.44	0.12
2	0.29	0.23
3	0.16	0.28
4 (spoiled)	0.12	0.14

TABLE IV: Grad-CAM mass-on-fruit by stage

Fig. 1: Grad-CAM saliency at the ripening endpoints. MobileNetV2 (top) concentrates on the fruit when fresh and disperses once spoiled; the more accurate MobileNetV3-Small (bottom) localises to the background at both stages.

Explore similar work

What Matters for Grocery Product Retrieval with Open Source Vision Language Models
May 18, 2026Emmanuel G. Maminta, Rowel O. AtienzaRecent Vision-Language Models Multimodal Retrieval
AI Graduate Program, University of the Philippines, Diliman, Quezon City · EEEI, University of the Philippines, Diliman, Quezon City
Low-Rank Adaptation of Frozen Vision-Language Models for Blind Image Quality Assessment
Jun 1, 2026Bishr Omer Adam, Xu LiImage Quality Assessment Frozen Vision-Language Models
School of Electronics and Information, Northwestern Polytechnical University, Xi’an, China
Few-Shot Ordinal Learning for Day-Wise Freshness Estimation with Hyperspectral Fish Images
Aug 12, 2026Kazi Nabiul Alam, Pooneh Bagheri Zadeh, Akbar Sheikh-AkbariHyperspectral Image Few-Shot Learning
School of Built Environment, Engineering and Computing, Leeds Beckett University, Leeds, United Kingdom.

May 18, 2026, cs.CVJ/K move · Enter open · S save

What Matters for Grocery Product Retrieval with Open Source Vision Language Models

Emmanuel G. Maminta, Rowel O. Atienza

AI Graduate Program, University of the Philippines, Diliman, Quezon City · EEEI, University of the Philippines, Diliman, Quezon City

Recent Vision-Language Models Multimodal Retrieval

Open paper AskPDF arXiv

Multimodal product retrieval (MPR) underpins checkout-free retail and automated inventory systems, yet it demands fine-grained SKU discrimination that standard vision-language benchmarks fail to capture. We present the first systematic zero-shot evaluation of 190 open-source VLMs on the MPR task of the GroceryVision Challenge, isolating pre-training data, architecture, and input resolution. Our analysis yields three actionable findings. \textbf{(1) Data quality trumps scale.} Switching from raw web-scrapes to filtered datasets delivers up to 16.6% accuracy gains, exceeding the benefit of doubling model parameters. \textbf{(2) Efficient models can win.} MobileCLIP-B (150M parameters) outperforms 351M counterparts trained on noisy data. We introduce \textit{semantic power density} (

φ

), an efficiency metric that penalizes sub-threshold accuracy. \textbf{(3) A precision gap persists.} State-of-the-art models achieve 94.5% Recall@5 but suffer a 17.5% drop at Recall@1, revealing that contrastive embeddings cluster categories effectively but fail to rank visually similar SKUs. Code and evaluation scripts are available at https://github.com/upeee/openmpr.

Abstract

Explore similar work

What Matters for Grocery Product Retrieval with Open Source Vision Language Models

Low-Rank Adaptation of Frozen Vision-Language Models for Blind Image Quality Assessment

Few-Shot Ordinal Learning for Day-Wise Freshness Estimation with Hyperspectral Fish Images