This project implements real-time portrait segmentation using semantic segmentation to separate people from backgrounds in images. It experiments with multiple encoder-decoder architectures including Mobile-Unet, DeeplabV3+, and SINet, training models on portrait datasets and comparing performance with standard evaluation metrics. The trained models are then optimized and deployed to edge devices (Jetson TX2, Coral TPU) using TensorFlow Lite and other mobile ML platforms for real-time inference.