Vanilla U-Net

Vanilla U-Net is the standard encoder-decoder convolutional network pattern for image segmentation. Introduced by Olaf Ronneberger, Philipp Fischer, and Thomas Brox at the University of Freiburg in Germany, it was first presented for biomedical images but became a general segmentation baseline. The contracting path captures context; the expanding path restores resolution, and skip connections copy fine detail across the U-shaped network. ‘Vanilla’ means the basic architecture, not a flavor-specific model or VGG19 variant. TT-Metal’s matrix uses 480x640 inputs and lists batch 1 on n150 Wormhole or batch 2 on n300 Wormhole. Those two rows are single- and two-device scales for the same segmentation architecture.