如何处理点云?主要可以分为两大类:
geometric information might be lost due to quantization, and voxels typically bring extra memory and computational costs
-
Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015.
-
SPLATNet: Sparse lattice networks for point cloud processing. In CVPR, 2018.
-
Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017.
MLP系列
使用的都是shared mlp,这限制了表达不同空间信息关系的能力
they all adopt the shared MLPs to transform point features, which limits the model capabilities in capturing spatial-variant information.
-
PointNet
-
PointNet2
-
3d graph neural networks for rgbd semantic segmentation. In ICCV, 2017.
-
Local spectral graph convolution for point set feature learnin. In ECCV,2018
-
A-cnn: Annularly convolutional neural networks on point clouds. In CVPR, 2019.
-
Dynamic graph cnn for learning on point clouds. ACM Trans. Graph., 2019.
卷积系列:探索局部邻域内空间关系
基于相对空间信息直接预测卷积核权重
缺陷:much higher complexity (memory and computation)
-
Spidercnn:Deep learning on point sets with parameterized convolutional filters. In ECCV, 2018
-
Deep parametric continuous convolutional neural networks. In CVPR, 2018
-
RS-CNN . In CVPR 2019
-
Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019
卷积的另一个方向就是使用固定的卷积点,通过计算点之间的相关度来调整卷积核的权重 limit the model flexibility
-
Point convolutional neural networks by extension operators. ACM Trans. Graph., 2018.
-
Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019.
-
Interpoated convolutional networks for 3d point cloud understanding. In ICCV, 2019
-
Seggcn: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020
动态卷积、条件卷积。Dynamic and conditioned convolutions.
本文与Condconv较为相识,但是Condconv预测的卷积核不是根据位置来的,但是在无序点云中需要根据空间位置来确定权重
-
Dynamic filter networks. In NeurIPS, 2016.
-
long-range interactions without attention ,2021
-
Condconv: Conditionally parameterized convolutions for efficient inference. In NeurIPS, 2019. 3