Aerial building segmentation (U-Net, ResNet34)
U-Net with an ImageNet-pretrained ResNet34 encoder (segmentation-models-pytorch), trained to segment buildings
in aerial images at ~1 m/pixel. Weights only (best.pth, state dict); build the model with
smp.Unet("resnet34", encoder_weights=None, in_channels=3, classes=1) and apply a sigmoid to the logits.
- Data: Massachusetts Buildings dataset (137 train / 4 val / 10 test images, 1500x1500).
- Loss: BCE + Dice, 40 epochs x 400 random 512x512 crops, AdamW, one seed.
- Test (full images, tiled inference): IoU 0.695, Dice 0.820 (Focal+Dice: 0.690 / 0.817, effectively a tie).
- Limitations: one region only, no cross-city evaluation; 1 m/pixel imagery; single image, no change detection.
Code, training and failure analysis: https://github.com/ArthDubey1011/aerial-building-segmentation