unet-fun 2024

projects

unet-fun demo — U-Net for segmentation: an image flows down the encoder, through the bottleneck, and up the decoder to a mask, with skip connections restoring spatial detail.

A from-scratch U-Net implemented in PyTorch for image segmentation, applied to the Carvana car-masking task — predicting a per-pixel mask that separates the car from its background.

The architecture

U-Net is an encoder–decoder with skip connections (shown in the animation above):

Encoder — repeated DoubleConv blocks (two Conv → BatchNorm → ReLU layers) followed by max-pooling. Spatial resolution halves at each step while the channel count doubles (64 → 128 → 256 → 512).
Bottleneck — the lowest-resolution, richest-feature representation.
Decoder — transposed convolutions upsample back toward the input resolution; at each level the matching encoder feature map is concatenated via a skip connection, so fine spatial detail lost during downsampling is restored.
Head — a final 1×1 convolution maps to a single channel; a sigmoid turns the logits into a binary foreground mask.

Training

The model is trained with binary cross-entropy on the Carvana images and masks, with albumentations augmentation and accuracy / Dice checks on a validation split.