Semantic Segmentation

Implemented an encoder–decoder segmentation network in PyTorch — a strided-convolution encoder (64 → 128 → 256 → 512 channels) with batch normalization and an upsampling decoder that reconstructs per-pixel class maps across 36 classes.

Trained pixel-wise classification with cross-entropy loss and built both baseline and improved variants of the architecture.

Created the dataset-preparation and Kaggle submission tooling used to evaluate predictions.