The algorithms we considered

• Predator https://github.com/prs-eth/OverlapPredator
• 3DMatch: https://github.com/andyzeng/3dmatch-toolbox
• DeepVCP (it is a local registration algorithm)
• 3DFeat-Net: https://github.com/yewzijian/3DFeatNet
• PPFNet: non official implementation
• Feature Metric Registration: https://github.com/XiaoshuiHuang/fmr
• DCP: https://github.com/WangYueFt/dcp
• 3DSmoothNet: https://github.com/zgojcic/3DSmoothNet
• GenReg: non official implementation
• Fully-convolutional geometric feature: https://github.com/chrischoy/FCGF
• RGM: https://github.com/fukexue/RGM
• RPM-Net: https://github.com/yewzijian/RPMNet
• DeepGRM: ?
• PointGMM: https://github.com/amirhertz/pointgmm
• CorsNet: no implementation
• RoReg: https://github.com/HpWang-whu/RoReg
• GeoTransformer: https://github.com/qinzheng93/GeoTransformer

Non end-to-end approaches are tested using the L2 distance and a kdtree to look for correspondences between features, and TEASER++, Fast Global Registration, or RANSAC to estimate the transformation from the correspondences.