Point cloud for classification

1 minute read

State of the art networks for different tasks

From the above link, it is observed that the most pupular datasets for point cloud classification is MOdelNet40 and ScanObjectNN.

For MOdelNet40 (released in 2015), in the top 10 networks, I first exclude the networks with extra training dataset, I can obtain:

They are all published in 2021 or 2022.

For ScanObjectNN (released in 2019), I first exclude the networks with extra training dataset, I can obtain:

They are all published in 2022.

Why the top networks for two datasets are so different? Which network should I choose for my dataset (PFT regression from binary vessel tree)?

ModelNet40 is synthetic, while ScanObjectNN is real-world dataset.

So I prefer to try the top networks in ScanObjectNN, which includes:

PointNet. Processed raw point sets through Multi-Layer Perceptrons (MLPs). While aggregating features at the global level using max-pooling operation, lost valuable local geometric information.
PointNet++. employed ball querying and k-Nearest Neighbor (k-NN) querying to query local neighborhoods to extract local semantic information. But it still lost contextual information due to the max-pooling operation
PointNeXt.
- Data augmentation
- we find that neither appending more SA blocks nor using more channels leads to a noticeable improvement in accuracy, while causing a significant drop in throughput
PointNeXt+Local. In previous paper, once the local features are obtained, the original neighborhood points, the directional vectors, and the distance computed (in the case of ball querying) are discarded. In this paper, we use the radius-normalized distance and directional vectors as additional local neighborhood features with minimal additional memory or computational costs.