Thesis reading: geotransformer

Abstract ：

Put forward geometric transformer Learning geometric features is robust superpoint matching . It encodes pair-wise Distance and triplet-wise angle , Make it robust to low repetition rate and rigid transformation invariance . The simple self attention mechanism achieves high matching accuracy without RANSAC, And the accuracy is improved 100 times . The internal point rate has increased by% 17-30 A little bit , The recall rate has increased 7 There's a point in 3DLoMatch.

Introduce

Inspired by image matching , Methods without key points , By downsampling superpoint Then match by checking whether their neighbors repeat , Thus, dense matching .

classification ：

Method based on matching points ：1、 Extract key points and descriptors for matching .2、 Consider all possible matches without the need for key points （ The method of this paper ）

Direct registration method ： End to end networks 1、 be based on ICP and SVD Decomposed thoughts 2、 Extract global features for registration . It is good for artificial shapes, but it may fail for large scenes .

Robust depth estimation ： Classification network eliminates outliers . shortcoming ： Need to train the model

Method ：

Problem model ：

C* Is the point on the match .C* Unknown , You need to estimate the matching relationship first .

This paper adopts a hierarchical matching method , From thick to thin , Simultaneous adoption KPConv-FPN Downsampling input dense point clouds and extraction point-wise features . One Point Matching Module Used to match superpoint And dense points . Finally, a rigid transformation is estimated without RANSAC.

Select the closest point in the dense point cloud to the downsampling