multimodalfactbullishDescriptor-free matching naturally enables multi-detector training as heterogeneous keypoints can be optimized in shared geometry-only space without aligning descriptor spacesComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02486v1