AgroPest-12 Insect Detection and Classification Benchmark

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

Problem / goal

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

My role

Computer Vision Developer on a group benchmark project.

What I personally contributed

• Implemented and compared HOG/SIFT-BoVW/SVM, YOLOv8 and EfficientNet-B0, YOLOv11n, YOLOv12n, and Faster R-CNN with ResNet50-FPN.
• Evaluated fixed train, validation, and test splits with mAP@50, mAP@50:95, accuracy, precision, recall, F1-score, ROC-AUC, latency, and FPS.
• Analyzed failure cases involving small, low-contrast, and camouflaged insects to explain model tradeoffs beyond headline metrics.

Technical approach

• Established comparable data splits and evaluation criteria across classical feature pipelines, hybrid classifiers, and modern detectors.
• Measured detection, image-level classification, and runtime behavior with task-appropriate metrics.
• Used qualitative error review to connect false detections and missed insects to scale, contrast, and camouflage.

Key features

• Classical, hybrid, and deep learning model comparison.
• Detection and image-level classification evaluation.
• Latency and FPS analysis alongside accuracy metrics.
• Failure analysis for difficult insect imagery.

Impact / results

• YOLOv11n achieved 80.77% mAP@50 for detection and 97.73% image-level classification accuracy on the test set.
• Documented accuracy, runtime, and failure-mode tradeoffs across the evaluated model families.

What I learned

• Strong aggregate metrics can hide systematic failures on small or camouflaged objects.
• A benchmark is more useful when accuracy, speed, and qualitative failure modes are interpreted together.