AgroPest-12 Insect Detection and Classification Benchmark

A group computer vision benchmark evaluating insect detection and classification methods on the AgroPest-12 dataset.

Problem / goal

Agricultural pest monitoring needs models that can detect small, low-contrast, and camouflaged insects while remaining practical enough for timely use.

My role

Collaborator on a group computer vision benchmark.

What I personally contributed

• Compared classical, hybrid, and deep learning approaches for 12-class insect detection and classification.
• Evaluated models using mAP@50, COCO mAP, accuracy, precision, recall, F1-score, ROC-AUC, latency, and FPS.
• Analyzed failure cases involving small, low-contrast, and camouflaged insects.

Technical approach

• Compared HOG + SIFT-BoVW + SVM, YOLOv8 + EfficientNet-B0, YOLOv11n, YOLOv12n, and Faster R-CNN with ResNet-50 FPN.
• Evaluated detection, classification, and runtime performance using mAP@50, COCO mAP, accuracy, precision, recall, F1-score, ROC-AUC, latency, and FPS.
• Analysed failure cases involving small, low-contrast, and camouflaged insects.

Key features

• Classical and deep learning model comparison.
• Detection and image-level classification evaluation.
• Runtime analysis using latency and FPS.
• Failure-case analysis for difficult insect imagery.

Impact / results

• YOLOv11n achieved the strongest overall result with 80.77% mAP@50 and 97.73% image-level classification accuracy on the test set.
• Documented tradeoffs between classical methods, hybrid pipelines, and modern detectors.

What I learned

• Strong aggregate metrics can hide systematic failures on small or camouflaged objects.
• Benchmarking is most useful when accuracy, runtime, and failure modes are interpreted together.