My solution combines multi-modal data fusion and robust training techniques to tackle the solar panel counting challenge. Key components include:

1. Model ArchitectureBackbone : EfficientNetV2 variant for image feature extraction Metadata Integration : Encoded image origin (D/G) and placement type (roof/ground) via one-hot/dense embeddings Fusion : Concatenated visual features + metadata processed through a 2-layer regression head
2. Data StrategyCross-Validation : Stratified K-Fold to handle class imbalance Augmentation Pipeline :Dynamic spatial transforms (geometric + color) Targeted dropout patterns to reduce overfitting
3. Training ProtocolLoss : MAE-focused objective with gradient scaling Optimization : AdamW with cosine LR scheduling Infrastructure : Mixed-precision training for efficiency
4. Inference Enhancements : Test-time augmentation (TTA) with consistent preprocessing Prediction aggregation from multiple model checkpoints

Validation Insights

• Achieved steady MAE improvement across epochs (1.25-2.35 range)
• Metadata integration provided ~8% performance boost vs image-only baseline

This approach balances model capacity, data diversity, and regularization to handle the dataset's unique challenges. Would love to hear about others' strategies for metadata utilization and augmentation design!