agri-cv-evaluation

Agricultural CV Evaluation

Description

Comprehensive evaluation framework for agricultural computer vision covering standard metrics, robustness testing, edge device benchmarking, and ablation experiment automation.

When to Use

• Need to evaluate agricultural CV model performance
• Need to run ablation experiments
• Need to test model robustness under different conditions (lighting, weather, occlusion, noise)
• Need to generate publication-quality evaluation tables and charts
• Need to benchmark model efficiency (FPS, FLOPs)

Evaluation Dimensions

Standard Metrics

• Classification: Accuracy, Precision, Recall, F1-score (per-class & macro)
• Detection: mAP@50, mAP@50:95, Precision, Recall
• Segmentation: mIoU, Dice coefficient, Pixel Accuracy
• General: FLOPs, Parameters, FPS

Robustness Testing

Tests model accuracy degradation under environmental perturbations:

• Lighting: brightness ±50%, contrast ±40%
• Weather: fog, rain simulations
• Occlusion: 10%, 20%, 30%, 50% random occlusion
• Noise: Gaussian noise σ=0.05, 0.1, 0.2
• Resolution: 100%, 50%, 25% of original

Returns nested dict: {condition: {severity: accuracy}}

Ablation Studies

Systematically vary hyperparameters and collect metrics:

• Supports unlimited ablation variables
• Auto-generates LaTeX tables and bar charts
• Configurable metric tracking

Code Examples

Classification Metrics

from agri_cv_research.evaluation import compute_classification_metrics

metrics = compute_classification_metrics(
    y_true=[0, 1, 2, 0, 1],
    y_pred=[0, 1, 2, 0, 2],
    class_names=["healthy", "early_blight", "late_blight"]
)
# Returns: accuracy, f1_macro, f1_weighted, precision_macro, recall_macro,
#          per_class_f1, confusion_matrix
print(f"F1-macro: {metrics['f1_macro']:.4f}")
print(f"Accuracy: {metrics['accuracy']:.4f}")

Detection Metrics

from agri_cv_research.evaluation import compute_detection_metrics

pred_boxes = [
    {"boxes": [[10, 20, 50, 60]], "scores": [0.95], "labels": [0]},
    {"boxes": [[100, 110, 150, 160]], "scores": [0.88], "labels": [1]},
]
gt_boxes = [
    {"boxes": [[10, 20, 50, 60]], "labels": [0]},
    {"boxes": [[100, 110, 150, 160]], "labels": [1]},
]

metrics = compute_detection_metrics(pred_boxes, gt_boxes)
# Returns: mAP@50, mAP@50:95, precision, recall
print(f"mAP@50: {metrics['mAP@50']:.4f}")