🚀 Introducing LiDeRe: a lightweight readout that turns frozen vision backbones (e.g. DINOv3) into strong dense predictors. A tiny head → semantic segmentation, detection, pose, contour. Often beats SoTA methods with a fraction of the trainable parameters. At #CVPR2026, poster session 1 on Friday.

📣 Paper alert: We present dense attentive probing (DeAP), a method to measure the representation quality of various vision backbones for dense prediction tasks. It uses small, parameter-efficient readouts with learnable masks to generate dense predictions from backbone features of any size.