• 3D datasets are typically much smaller in quantity, with relatively
  small amount of labels and limited diversity
• data augmentation (DA) is a very common strategy to avoid overfitting and improve the network generalization ability by artificially enlarging the quantity and diversity of the training samples
• A bad projection function can easily lead to the loss of structural information in a point cloud with, for instance, many point collisions in the image space.
• Point clouds can be treated as graphs by associating edges among neighbors. This paves the way to the appliance of graph convolutional networks
• 3D datasets are typically much smaller in quantity, with relatively small amount of labels and limited diversity

《Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs》(CVPR 2017)

We identify that the current formulations of graph convolution do not exploit edge labels, which results in an overly homogeneous view of local graph neighborhoods。

• Table 4 conveys that while there is no clear winning algorithm, our method performs at the level of state of the art for edge-labeled datasets

《Image Transformer》(ICML 2019)

One disadvantage of CNNs compared to RNNs is their typically fairly limited receptive field. This can adversely affect their ability to model long-range phenomena common in images, such as symmetry and occlusion, especially with a small number of layers.

• we propose eschewing recurrent and convolutional networks in favor of the Image Transformer
• There is a broad variety of types of image generation models in the literature.
• While very promising, GANs have various drawbacks. They are notoriously unstable

《Relation Networks for Object Detection》(CVPR 2018)

• It is built upon a basic attention module.
• As a result, it serves as a basic building block that is usable in any architecture flexibly

《A Hierarchical Graph Network for 3D Object Detection on Point Clouds》(CVPR 2020)

3D object detection on point clouds finds many applications. However, most known point cloud object detection methods did not adequately accommodate the characteristics (e.g., sparsity) of point clouds

• But, there are still some challenging drawbacks.
• First, using PointNet++ as backbone neglected some local shape information, since the relative geometric positions of points were not accounted for. Second, the multi-level semantics were not adequately utilized by the structures of the frameworks, which might neglect some helpful information for object detection.
• In our framework, the local shape information, semantics of multilevels,
  and global scene information (features of proposals) of point clouds are sufficiently captured,** aggregated**, and incorporated by the hierarchical graph model, giving full consideration of the characteristics of point cloud data.
• Even though the higher-level features are computed by fusing the lower-level features in the upsampling pathway, it is more beneficial to use the multilevel features together for proposal generation as the features of different levels provide various semantics

《AM-GCN: Adaptive Multi-channel Graph Convolutional Networks》

• The weakness may severely hinder the capability of GCNs in some classification tasks, since GCNs may not be able to adaptively learn some deep correlation information between topological structures and node features.
• The results meet the expectation.