D3Former: Jointly learning repeatable dense detectors and feature-enhanced descriptors via saliency-guided transformer

Junjie Gao, Pengfei Wang, Qiujie Dong, Qiong Zeng, Shiqing Xin, Caiming Zhang

Abstract: Highlights•Saliency-guided transformer for point cloud registration, integrating dense detector and feature descriptor learning.•The feature enhancement module boosts feature discrimination via region attention, crucial for detailing texture-less areas.•The repeatable keypoints module boosts keypoint repeatability, crucial for robust matching.•Extensive experiments show D3Former consistently outperforms state-of-the-art methods across various benchmarks.