SAAIL Lab SAAIL Lab
CV4A @ CVPR 2026 Accepted & Published · CVPR 2026 Workshop

CLOVES-4603: Benchmarking Classical Texture Features and Fine-Tuned Deep Models for Clove Quality Grading

1School of Engineering and Science, Indian Institute of Technology Madras Zanzibar, Tanzania
2Wadhwani School of Data Science and AI (WSDSAI), Indian Institute of Technology Madras, Chennai, India
CVPR 2026 , 7th Workshop on Vision for Agriculture (CV4A / V4A), Denver, USA (May 2026)
CLOVES-4603: Benchmarking Classical Texture Features and Fine-Tuned Deep Models for Clove Quality Grading

Summary: Five-paradigm benchmark on CLOVES-4603 (4,603 ZSTC clove images): fine-tuned ResNet-50 reaches 99.67% accuracy, frozen DINOv2 linear probe achieves 94.90%, and GPU-free GLCM+GLRLM+LBP texture fusion SVM reaches 92.94%.

Abstract

We introduce CLOVES-4603, the first public benchmark dataset for automated clove quality grading, comprising 4,603 standardized 224x224 RGB images spanning four commercial grades collected at the Zanzibar State Trading Corporation (ZSTC). We benchmark classical texture pipelines (HOG+SVM, GLCM+GLRLM+LBP fusion) against fine-tuned deep learning models including ResNet-50, MobileNetV3, and a frozen DINOv2 vision transformer. ResNet-50 achieves 99.67% accuracy on the test set, whereas texture-fusion SVM reaches 92.94%, providing a strong and deployment-friendly classical baseline. This benchmark informs real-world deployment trade-offs between accuracy, latency, and computational cost in agricultural settings where infrastructure is limited.

Keywords: CLOVES-4603clove quality gradingZanzibartexture featuresdeep learningCVPR 2026vision for agriculture

Key Findings & Contributions

4,603
Curated Images
224x224 RGB · 4 ZSTC grades
99.67%
ResNet-50 Acc
99.66% Macro-F1 (n=921 test)
94.90%
DINOv2 Probe
Frozen ViT-B/14 baseline
92.94%
Texture-Fusion SVM
GLCM + GLRLM + LBP
  1. Introduces CLOVES-4603, the first open benchmark dataset of 4,603 lot-aware stratified images spanning four ISO 2254-2004 / ZSTC commercial clove quality grades collected in Zanzibar.
  2. Establishes rigorous 5-seed evaluation protocols with bootstrap 95% confidence intervals comparing classical texture descriptors, self-supervised vision transformers, and fine-tuned CNNs.
  3. Demonstrates that combined GLCM+GLRLM+LBP texture fusion with RBF-SVM achieves 92.94% test accuracy without GPUs, outperforming HOG+SVM (32.68%) by +60.26 percentage points.
  4. Profiles accuracy-latency-parameter Pareto trade-offs showing MobileNetV3-Large reaches 99.35% accuracy as a mobile-ready alternative to ResNet-50 (99.67%).

Key Empirical Figure

Lot-aware stratified 70/10/20 class distribution across Grades 1 to 4 (N = 4,603 ZSTC clove images).
Lot-aware stratified 70/10/20 class distribution across Grades 1 to 4 (N = 4,603 ZSTC clove images).

Benchmark Results

CLOVES-4603 Five-Paradigm Test Benchmark (n = 921 Held-Out Images)

Method / ArchitectureTest Accuracy (%)Training Mode / Features
ResNet-5099.67%Fine-tuned ImageNet CNN (99.66% F1)
MobileNetV3-Large99.35%Lightweight edge-deployable CNN
DINOv2 ViT-B/1494.90%Frozen self-supervised linear probe
GLCM + GLRLM + LBP SVM92.94%Classical 96-dim texture fusion
HOG + SVM32.68%Shape-gradient classical baseline

BibTeX

@inproceedings{nyalala2026cloves4603,
  title={CLOVES-4603: Benchmarking Classical Texture Features and Fine-Tuned Deep Models for Clove Quality Grading},
  author={Nyalala, Innocent and Vincent, Patrick Ndowo},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops -- Vision for Agriculture (CV4A)},
  year={2026},
  url={https://openreview.net/pdf?id=V1UbAtXS4M}
}
APA Citation

Nyalala, I., & Vincent, P. N. (2026). CLOVES-4603: Benchmarking Classical Texture Features and Fine-Tuned Deep Models for Clove Quality Grading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (V4A). https://openreview.net/pdf?id=V1UbAtXS4M