RIPE++: Reinforced Keypoint Learning from Positive Pairs Only

zhenjun_zhao · x · 2026-08-21

The paper presents RIPE++, a method for reinforced keypoint learning using only positive pairs. Unlike traditional RL approaches like RIPE that rely on binary rewards and negative pairs, RIPE++ exploits geometric consistency signals to derive both reward and penalty from a single positive pair. This rich signal enables learning discriminative detectors and descriptors without negative contrasting. The method also extends to the matching stage, improving AUC@5 on MegaDepth1500 from 56.58 to 59.65.

Original post →

More from Research

Research channel →