Implemented and compared three convolutional encoder-decoder architectures (U-Net from scratch, ResNet34-UNet, and fine-tuned DeepLabV3-ResNet50) for multi-class dense pixel predictions on Pascal VOC 2012 (21 classes)., Formulated a hybrid loss function combining Cross-Entropy and Dice Loss to counter extreme class imbalance and improve mean Intersection-over-Union (mIoU) under proper ignore-index handling., Built automated benchmarking scripts evaluating per-class IoU and pixel-wise confusion matrices under CUDA acceleration, plus an interactive inference app for real-time segmentation visualization.