- 3D Point Cloud
3D Point Cloud
3D shape classification 3D object detection and tracking 3D point cloud segmentation
3D shape classification
PointNet
3D object detection and tracking
1. Region Proposal-based :
- 提出几个可能的包含对象的区域，然后提取区域特征以确定每个提议的类别标签。
multi-view based: 慢
从 BEV 图生成一组高度准确的 3D 候选框，并将它们投影到多个视图的特征图（例如，LiDAR 前视图图像、RGB 图像）。然后他们结合来自不同视图的这些区域特征来预测定向的 3D 边界框.
RT3D使用了预 RoI 池化卷积来提高MV3D 的效率。具体来说，他们将大部分卷积操作移到了 RoI 池化模块之前。因此，对所有对象建议执行一次 RoI 卷积。实验结果表明，该方法可以以 11.1 fps 的速度运行，比 MV3D 快 5 倍。
segmentation-based：STD（F）/PointPainting/PointRGCN
- 语义分割技术去除大部分背景点，在前景点上生成大量高质量的提议以节省计算量，如图 8(b) 所示。与multi-view based相比，这些方法实现了更高的对象召回率，更适用于具有高度遮挡和拥挤对象的复杂场景。
frustum-based : F-ConvNet
- leverage existing 2D object detectors to generate 2D candidate regions of objects and then extract a 3D frustum proposal for each 2D candidate region. 受到 2D object detectors的限制。
2. Single Shot Methods :
- directly predict class probabilities and regress 3D bounding boxes of objects using a single-stage network. They do not need region proposal generation and post-processing.
- fast
BEV-based
Discretization-based : 3DBN/SA-SSD/
- convert a point cloud into a regular discrete representation, and then apply CNN to predict both categories and 3D boxes of objects.
Point-based Methods : 3DSSD
- 直接以原始点云为输入
other ： LaserNet/Lasernet++
3D point cloud segmentation
Grid-GCNN
LDG CNN
PointWeb
- detection
- acc not good
pointconv
1000 points<9.9ms
PAConv
batch size = 16 0.118s / batch 7ms / 1w points acc : 0.935981
