EfficientNetV2B0 RGB + CNN Depth Carbohydrate Regression

Architecture

• Dual-input multimodal regression model
• One pretrained EfficientNetV2B0 branch for overhead RGB images
• One from-scratch CNN branch for overhead depth images
• Global average pooling on both branches
• Feature fusion through concatenation
• Fully connected regression head
• Final dense layer with linear activation for carbohydrate prediction

Backbone setup

• EfficientNetV2B0 is initialized with ImageNet pretrained weights for RGB
• The RGB backbone is frozen during the initial training stage
• The depth branch is trained from scratch

Input modalities

• rgb_input: overhead RGB image
• depth_input: overhead depth image

Target

• total_carb