This project explores semantic segmentation of buildings from aerial imagery using PyTorch and the DeepLabV3+ architecture (with a ResNet50 backbone). The primary focus is on improving boundary prediction accuracy through the use of custom composite loss functions, particularly incorporating edge-weighted components.
- Model: DeepLabV3+ with ResNet50 backbone (using
torchvision). - Datasets: Primarily uses the Massachusetts Buildings Dataset (support for others like Inria can be adapted). Data is handled externally via Google Drive (see
data/README.md). - Loss Functions: Implements a composite loss combining Binary Cross-Entropy (BCE), Tversky loss, and an adapted Hinge loss.
- Edge Weighting: Explores applying Sobel-filter-derived edge weights to loss components (specifically Hinge loss) to emphasize boundary learning.
- Refinement Layer: Includes an option to add a final 3x3 convolutional layer for potential output refinement.
- Comparative Experiments: Scripts designed to train two models in parallel with identical initial weights but different loss configurations for direct comparison.
- Visualization: Generates detailed visualizations comparing model outputs (probability maps, error maps, loss component maps) over epochs.
- Animation: Creates animated GIFs showing the evolution of predictions and metrics over training epochs.
- Metrics: Comprehensive evaluation metrics (IoU, Dice, Precision, Recall, TP/FP/FN/TN) logged during training and validation.