Plant Pathology CVPR 2020

This competition is part of the Fine-Grained Visual Categorization FGVC7 workshop at the Computer Vision and Pattern Recognition Conference CVPR 2020.

Dataset Overview:

class	images
healthy	516
multiple_diseases	91
rust	622
scab	592
	Total: 1821

Solutions:

Data Preprocessing and Augmentations:

• Resize(256, 256)
• CenterCrop(224, 224)
• HorizontalFlip(), VerticalFlip()
• ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.1, rotate_limit=15)
• RandomBrightness()
• HueSaturationValue()
• Normalize()

Models

• ResNet50
• SE ResNet50

Training Settings

• Validation Size = 0.1
• Batch Size = 8
• Epochs = 40
• Loss function: Class-Balanced Binary Cross-Entropy Loss (beta = 0.99)
• Regularization: Weight Decay (lambda = 0.001)
• Optimizer: SGD (lr=0.01, momentum=0.9)
• Learning Rate Scheduler: ReduceLROnPlateau (factor=0.1, patience=5)

Results

	ResNet50	SE ResNet50
Train Accuracy	98.17 %	99.63 %
Valid Accuracy	96.72 %	97.81 %
Train Confusion Matrix
Valid Confusion Matrix
Public Score	0.942	0.947