Text this: Learn Robust Pedestrian Representation Within Minimal Modality Discrepancy for Visible-Infrared Person Re-Identification.