AIJonas/nutrition5k-baseline-cnn-rgb-depth-side-carb

Baseline CNN RGB + Depth + Side-angle Carbohydrate Regression

This model predicts dish-level carbohydrate content from:

• overhead RGB image
• overhead depth image
• one side-angle RGB image at a time

All available side-angle images are used during training by expanding each dish into multiple samples.

Dataset

Based on Nutrition5K compact preprocessing.

Split counts (rows after side-view expansion)

• Train rows: 17348
• Validation rows: 2652
• Test rows: 3596

Unique dishes

• Train dishes: 2303
• Validation dishes: 334
• Test dishes: 474

Architecture

• RGB overhead CNN branch
• Depth overhead CNN branch
• Side-angle RGB CNN branch
• Feature fusion via concatenation
• Dense regression head
• Target: total_carb

Training

• Loss: MSE
• Metrics: MAE, MSE
• Optimizer: Adam
• Best checkpoint selected by validation loss

Notes

Each available side-angle image becomes a separate sample while keeping the same overhead RGB and depth image for the corresponding dish.