UFLD-v2

UFLD-v2, short for Ultra Fast Lane Detection v2, is a specialized computer-vision model for finding road-lane geometry. Researchers at Zhejiang University in China formulate lane finding as hybrid-anchor-driven ordinal classification: the model predicts ordered lane coordinates along row and column anchors instead of producing a dense pixel mask like an ordinary semantic segmenter. That task-specific representation is its memory hook and explains why ‘segmentation’ is only the matrix section, not the best description of its output. TT-Metal lists 320x800 inputs on n150 Wormhole at batch 1, n300 Wormhole at DP=2 and batch 2, and p150 Blackhole at batch 1. These three hardware rows collapse into one lane-detection architecture.