Paper reading: the perfect match: 3D point cloud matching with smoothed densities

Abstract

In this paper, 3DSmoothNet, Using a very similar deep learning architecture and full convolution , Use voxels to smooth density values （SDV） Express . The latter is calculated at the points of interest and arranged into local reference frames （LRF） To achieve rotation invariance .

94.9% The recall rate is 3DMatch On benchmark dataset . Output only 32 dimension , Such a low dimensional output is allowed in the standard PC In order to 0.1ms Search at the speed of each point . because SVD and LRF And high-quality descriptive features with full convolution learning , Our method is that sensors and scenes are not fixed .

Introduce

Two types of registration ： Based on descriptors and geometric features . Narrator ： Manually set to describe neighbors with histograms , Or local surface standard .

quite a lot 3D The feature learning network is neither rotation invariant , And it requires a very large output dimension to succeed , Or it is difficult to generalize to new fields .

The characteristics of this article ： There are low output dimensions （16 or 32）, High descriptor , Strong generalization ability , Rotation invariance .

SDV It reduces the sparsity of the input voxel mesh , It can get a better gradient in backward propagation , At the same time, reduce the influence of the boundary as well as smoothing out small miss-alignments due to errors in the estimation of the local reference frame (LRF)

Essentially unchanging feature ：point pair features

pointnet and pointnet++ Is the pioneer of a framework , They showed a network of transformation invariance , But they are not satisfactory in extracting geometric information , Because they cannot use volume grassroots in the network .

PPF use Pointnet But non rotational invariance ,PPF-foldnet Solved rotation invariance .

This paper establishes the relationship between the two point clouds as follows ：

If Q and P The match is successful , that K for 1, There are three unknowns in the formula , By mapping to higher dimensions, we can solve K, Thus the transformation matrix is solved .