PSMNet implements a pyramid stereo matching network that uses spatial pyramid pooling to aggregate global context at multiple scales and a 3D CNN with stacked hourglass networks to regularize cost volumes for disparity estimation. The model takes stereo image pairs as input and outputs per-pixel depth estimates, demonstrating that global context information helps resolve ambiguity in matching regions. This is a supervised learning approach for disparity prediction using convolutional neural networks trained on stereo datasets.